obtained gcc-4.6.4.tar.bz2 from upstream website;upstream

verified gcc-4.6.4.tar.bz2.sig;
imported gcc-4.6.4 source tree from verified upstream tarball.

downloading a git-generated archive based on the 'upstream' tag
should provide you with a source tree that is binary identical
to the one extracted from the above tarball.

if you have obtained the source via the command 'git clone',
however, do note that line-endings of files in your working
directory might differ from line-endings of the respective
files in the upstream repository.

12 files changed, 6348 insertions, 0 deletions

diff --git a/boehm-gc/include/private/cord_pos.h b/boehm-gc/include/private/cord_pos.h
new file mode 100644
index 000000000..d2b24bb8a
--- /dev/null
+++ b/boehm-gc/include/private/cord_pos.h
@@ -0,0 +1,118 @@
+/* 
+ * Copyright (c) 1993-1994 by Xerox Corporation.  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.
+ */
+/* Boehm, May 19, 1994 2:23 pm PDT */
+# ifndef CORD_POSITION_H
+
+/* The representation of CORD_position.  This is private to the	*/
+/* implementation, but the size is known to clients.  Also	*/
+/* the implementation of some exported macros relies on it.	*/
+/* Don't use anything defined here and not in cord.h.		*/
+
+# define MAX_DEPTH 48
+	/* The maximum depth of a balanced cord + 1.		*/
+	/* We don't let cords get deeper than MAX_DEPTH.	*/
+
+struct CORD_pe {
+    CORD pe_cord;
+    size_t pe_start_pos;
+};
+
+/* A structure describing an entry on the path from the root 	*/
+/* to current position.						*/
+typedef struct CORD_Pos {
+    size_t cur_pos;
+    int path_len;
+#	define CORD_POS_INVALID (0x55555555)
+		/* path_len == INVALID <==> position invalid */
+    const char *cur_leaf;	/* Current leaf, if it is a string.	*/
+    				/* If the current leaf is a function,	*/
+    				/* then this may point to function_buf	*/
+    				/* containing the next few characters.	*/
+    				/* Always points to a valid string	*/
+    				/* containing the current character 	*/
+    				/* unless cur_end is 0.			*/
+    size_t cur_start;	/* Start position of cur_leaf	*/
+    size_t cur_end;	/* Ending position of cur_leaf	*/
+    			/* 0 if cur_leaf is invalid.	*/
+    struct CORD_pe path[MAX_DEPTH + 1];
+    	/* path[path_len] is the leaf corresponding to cur_pos	*/
+    	/* path[0].pe_cord is the cord we point to.		*/
+#   define FUNCTION_BUF_SZ 8
+    char function_buf[FUNCTION_BUF_SZ];	/* Space for next few chars	*/
+    					/* from function node.		*/
+} CORD_pos[1];
+
+/* Extract the cord from a position:	*/
+CORD CORD_pos_to_cord(CORD_pos p);
+	
+/* Extract the current index from a position:	*/
+size_t CORD_pos_to_index(CORD_pos p);
+	
+/* Fetch the character located at the given position:	*/
+char CORD_pos_fetch(CORD_pos p);
+	
+/* Initialize the position to refer to the give cord and index.	*/
+/* Note that this is the most expensive function on positions:	*/
+void CORD_set_pos(CORD_pos p, CORD x, size_t i);
+	
+/* Advance the position to the next character.	*/
+/* P must be initialized and valid.		*/
+/* Invalidates p if past end:			*/
+void CORD_next(CORD_pos p);
+
+/* Move the position to the preceding character.	*/
+/* P must be initialized and valid.			*/
+/* Invalidates p if past beginning:			*/
+void CORD_prev(CORD_pos p);
+	
+/* Is the position valid, i.e. inside the cord?		*/
+int CORD_pos_valid(CORD_pos p);
+
+char CORD__pos_fetch(CORD_pos);
+void CORD__next(CORD_pos);
+void CORD__prev(CORD_pos);
+
+#define CORD_pos_fetch(p)	\
+    (((p)[0].cur_end != 0)? \
+     	(p)[0].cur_leaf[(p)[0].cur_pos - (p)[0].cur_start] \
+     	: CORD__pos_fetch(p))
+
+#define CORD_next(p)	\
+    (((p)[0].cur_pos + 1 < (p)[0].cur_end)? \
+    	(p)[0].cur_pos++ \
+    	: (CORD__next(p), 0))
+
+#define CORD_prev(p)	\
+    (((p)[0].cur_end != 0 && (p)[0].cur_pos > (p)[0].cur_start)? \
+    	(p)[0].cur_pos-- \
+    	: (CORD__prev(p), 0))
+
+#define CORD_pos_to_index(p) ((p)[0].cur_pos)
+
+#define CORD_pos_to_cord(p) ((p)[0].path[0].pe_cord)
+
+#define CORD_pos_valid(p) ((p)[0].path_len != CORD_POS_INVALID)
+
+/* Some grubby stuff for performance-critical friends:	*/
+#define CORD_pos_chars_left(p) ((long)((p)[0].cur_end) - (long)((p)[0].cur_pos))
+	/* Number of characters in cache.  <= 0 ==> none	*/
+
+#define CORD_pos_advance(p,n) ((p)[0].cur_pos += (n) - 1, CORD_next(p))
+	/* Advance position by n characters	*/
+	/* 0 < n < CORD_pos_chars_left(p)	*/
+
+#define CORD_pos_cur_char_addr(p) \
+	(p)[0].cur_leaf + ((p)[0].cur_pos - (p)[0].cur_start)
+	/* address of current character in cache.	*/
+
+#endif
diff --git a/boehm-gc/include/private/darwin_semaphore.h b/boehm-gc/include/private/darwin_semaphore.h
new file mode 100644
index 000000000..0f43982d5
--- /dev/null
+++ b/boehm-gc/include/private/darwin_semaphore.h
@@ -0,0 +1,68 @@
+#ifndef GC_DARWIN_SEMAPHORE_H
+#define GC_DARWIN_SEMAPHORE_H
+
+#if !defined(GC_DARWIN_THREADS)
+#error darwin_semaphore.h included with GC_DARWIN_THREADS not defined
+#endif
+
+/*
+   This is a very simple semaphore implementation for darwin. It
+   is implemented in terms of pthreads calls so it isn't async signal
+   safe. This isn't a problem because signals aren't used to
+   suspend threads on darwin.
+*/
+   
+typedef struct {
+    pthread_mutex_t mutex;
+    pthread_cond_t cond;
+    int value;
+} sem_t;
+
+static int sem_init(sem_t *sem, int pshared, int value) {
+    int ret;
+    if(pshared)
+        GC_abort("sem_init with pshared set");
+    sem->value = value;
+    
+    ret = pthread_mutex_init(&sem->mutex,NULL);
+    if(ret < 0) return -1;
+    ret = pthread_cond_init(&sem->cond,NULL);
+    if(ret < 0) return -1;
+    return 0;
+}
+
+static int sem_post(sem_t *sem) {
+    if(pthread_mutex_lock(&sem->mutex) < 0)
+        return -1;
+    sem->value++;
+    if(pthread_cond_signal(&sem->cond) < 0) {
+        pthread_mutex_unlock(&sem->mutex);
+        return -1;
+    }
+    if(pthread_mutex_unlock(&sem->mutex) < 0)
+        return -1;
+    return 0;
+}
+
+static int sem_wait(sem_t *sem) {
+    if(pthread_mutex_lock(&sem->mutex) < 0)
+        return -1;
+    while(sem->value == 0) {
+        pthread_cond_wait(&sem->cond,&sem->mutex);
+    }
+    sem->value--;
+    if(pthread_mutex_unlock(&sem->mutex) < 0)
+        return -1;    
+    return 0;
+}
+
+static int sem_destroy(sem_t *sem) {
+    int ret;
+    ret = pthread_cond_destroy(&sem->cond);
+    if(ret < 0) return -1;
+    ret = pthread_mutex_destroy(&sem->mutex);
+    if(ret < 0) return -1;
+    return 0;
+}
+
+#endif
diff --git a/boehm-gc/include/private/darwin_stop_world.h b/boehm-gc/include/private/darwin_stop_world.h
new file mode 100644
index 000000000..f6f5314ee
--- /dev/null
+++ b/boehm-gc/include/private/darwin_stop_world.h
@@ -0,0 +1,22 @@
+#ifndef GC_DARWIN_STOP_WORLD_H
+#define GC_DARWIN_STOP_WORLD_H
+
+#if !defined(GC_DARWIN_THREADS)
+#error darwin_stop_world.h included without GC_DARWIN_THREADS defined
+#endif
+
+#include <mach/mach.h>
+#include <mach/thread_act.h>
+
+struct thread_stop_info {
+    mach_port_t mach_thread;
+};
+
+struct GC_mach_thread {
+  thread_act_t thread;
+  int already_suspended;
+};
+
+void GC_darwin_register_mach_handler_thread(mach_port_t thread);
+
+#endif
diff --git a/boehm-gc/include/private/dbg_mlc.h b/boehm-gc/include/private/dbg_mlc.h
new file mode 100644
index 000000000..e0a994de5
--- /dev/null
+++ b/boehm-gc/include/private/dbg_mlc.h
@@ -0,0 +1,175 @@
+/* 
+ * Copyright 1988, 1989 Hans-J. Boehm, Alan J. Demers
+ * Copyright (c) 1991-1995 by Xerox Corporation.  All rights reserved.
+ * Copyright (c) 1997 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.
+ */
+
+/*
+ * This is mostly an internal header file.  Typical clients should
+ * not use it.  Clients that define their own object kinds with
+ * debugging allocators will probably want to include this, however.
+ * No attempt is made to keep the namespace clean.  This should not be
+ * included from header files that are frequently included by clients.
+ */
+
+#ifndef _DBG_MLC_H
+
+#define _DBG_MLC_H
+
+# define I_HIDE_POINTERS
+# include "gc_priv.h"
+# ifdef KEEP_BACK_PTRS
+#   include "gc_backptr.h"
+# endif
+
+#ifndef HIDE_POINTER
+  /* Gc.h was previously included, and hence the I_HIDE_POINTERS	*/
+  /* definition had no effect.  Repeat the gc.h definitions here to	*/
+  /* get them anyway.							*/
+    typedef GC_word GC_hidden_pointer;
+#   define HIDE_POINTER(p) (~(GC_hidden_pointer)(p))
+#   define REVEAL_POINTER(p) ((GC_PTR)(HIDE_POINTER(p)))
+#endif /* HIDE_POINTER */
+
+# define START_FLAG ((word)0xfedcedcb)
+# define END_FLAG ((word)0xbcdecdef)
+	/* Stored both one past the end of user object, and one before	*/
+	/* the end of the object as seen by the allocator.		*/
+
+# if defined(KEEP_BACK_PTRS) || defined(PRINT_BLACK_LIST) \
+     || defined(MAKE_BACK_GRAPH)
+    /* Pointer "source"s that aren't real locations.	*/
+    /* Used in oh_back_ptr fields and as "source"	*/
+    /* argument to some marking functions.		*/
+#	define NOT_MARKED (ptr_t)(0)
+#	define MARKED_FOR_FINALIZATION (ptr_t)(2)
+	    /* Object was marked because it is finalizable.	*/
+#	define MARKED_FROM_REGISTER (ptr_t)(4)
+	    /* Object was marked from a rgister.  Hence the	*/
+	    /* source of the reference doesn't have an address.	*/
+# endif /* KEEP_BACK_PTRS || PRINT_BLACK_LIST */
+
+/* Object header */
+typedef struct {
+#   if defined(KEEP_BACK_PTRS) || defined(MAKE_BACK_GRAPH)
+	/* We potentially keep two different kinds of back 	*/
+	/* pointers.  KEEP_BACK_PTRS stores a single back 	*/
+	/* pointer in each reachable object to allow reporting	*/
+	/* of why an object was retained.  MAKE_BACK_GRAPH	*/
+	/* builds a graph containing the inverse of all 	*/
+	/* "points-to" edges including those involving 		*/
+	/* objects that have just become unreachable. This	*/
+	/* allows detection of growing chains of unreachable	*/
+	/* objects.  It may be possible to eventually combine	*/
+	/* both, but for now we keep them separate.  Both	*/
+	/* kinds of back pointers are hidden using the 		*/
+	/* following macros.  In both cases, the plain version	*/
+	/* is constrained to have an least significant bit of 1,*/
+	/* to allow it to be distinguished from a free list 	*/
+	/* link.  This means the plain version must have an	*/
+	/* lsb of 0.						*/
+	/* Note that blocks dropped by black-listing will	*/
+	/* also have the lsb clear once debugging has		*/
+	/* started.						*/
+	/* We're careful never to overwrite a value with lsb 0.	*/
+#       if ALIGNMENT == 1
+	  /* Fudge back pointer to be even.  */
+#	  define HIDE_BACK_PTR(p) HIDE_POINTER(~1 & (GC_word)(p))
+#	else
+#	  define HIDE_BACK_PTR(p) HIDE_POINTER(p)
+#	endif
+	
+#       ifdef KEEP_BACK_PTRS
+	  GC_hidden_pointer oh_back_ptr;
+#	endif
+#	ifdef MAKE_BACK_GRAPH
+	  GC_hidden_pointer oh_bg_ptr;
+#	endif
+#	if defined(ALIGN_DOUBLE) && \
+	    (defined(KEEP_BACK_PTRS) != defined(MAKE_BACK_GRAPH))
+	  word oh_dummy;
+#	endif
+#   endif
+    GC_CONST char * oh_string;	/* object descriptor string	*/
+    word oh_int;		/* object descriptor integers	*/
+#   ifdef NEED_CALLINFO
+      struct callinfo oh_ci[NFRAMES];
+#   endif
+#   ifndef SHORT_DBG_HDRS
+      word oh_sz;			/* Original malloc arg.		*/
+      word oh_sf;			/* start flag */
+#   endif /* SHORT_DBG_HDRS */
+} oh;
+/* The size of the above structure is assumed not to dealign things,	*/
+/* and to be a multiple of the word length.				*/
+
+#ifdef SHORT_DBG_HDRS
+#   define DEBUG_BYTES (sizeof (oh))
+#   define UNCOLLECTABLE_DEBUG_BYTES DEBUG_BYTES
+#else
+    /* Add space for END_FLAG, but use any extra space that was already	*/
+    /* added to catch off-the-end pointers.				*/
+    /* For uncollectable objects, the extra byte is not added.		*/
+#   define UNCOLLECTABLE_DEBUG_BYTES (sizeof (oh) + sizeof (word))
+#   define DEBUG_BYTES (UNCOLLECTABLE_DEBUG_BYTES - EXTRA_BYTES)
+#endif
+
+/* Round bytes to words without adding extra byte at end.	*/
+#define SIMPLE_ROUNDED_UP_WORDS(n) BYTES_TO_WORDS((n) + WORDS_TO_BYTES(1) - 1)
+
+/* ADD_CALL_CHAIN stores a (partial) call chain into an object	*/
+/* header.  It may be called with or without the allocation 	*/
+/* lock.							*/
+/* PRINT_CALL_CHAIN prints the call chain stored in an object	*/
+/* to stderr.  It requires that we do not hold the lock.	*/
+#ifdef SAVE_CALL_CHAIN
+#   define ADD_CALL_CHAIN(base, ra) GC_save_callers(((oh *)(base)) -> oh_ci)
+#   define PRINT_CALL_CHAIN(base) GC_print_callers(((oh *)(base)) -> oh_ci)
+#else
+# ifdef GC_ADD_CALLER
+#   define ADD_CALL_CHAIN(base, ra) ((oh *)(base)) -> oh_ci[0].ci_pc = (ra)
+#   define PRINT_CALL_CHAIN(base) GC_print_callers(((oh *)(base)) -> oh_ci)
+# else
+#   define ADD_CALL_CHAIN(base, ra)
+#   define PRINT_CALL_CHAIN(base)
+# endif
+#endif
+
+# ifdef GC_ADD_CALLER
+#   define OPT_RA ra,
+# else
+#   define OPT_RA
+# endif
+
+
+/* Check whether object with base pointer p has debugging info	*/ 
+/* p is assumed to point to a legitimate object in our part	*/
+/* of the heap.							*/
+#ifdef SHORT_DBG_HDRS
+# define GC_has_other_debug_info(p) TRUE
+#else
+  GC_bool GC_has_other_debug_info(/* p */);
+#endif
+
+#if defined(KEEP_BACK_PTRS) || defined(MAKE_BACK_GRAPH)
+# define GC_HAS_DEBUG_INFO(p) \
+	((*((word *)p) & 1) && GC_has_other_debug_info(p))
+#else
+# define GC_HAS_DEBUG_INFO(p) GC_has_other_debug_info(p)
+#endif
+
+/* Store debugging info into p.  Return displaced pointer. */
+/* Assumes we don't hold allocation lock.		   */
+ptr_t GC_store_debug_info(/* p, sz, string, integer */);
+
+#endif /* _DBG_MLC_H */
diff --git a/boehm-gc/include/private/gc_hdrs.h b/boehm-gc/include/private/gc_hdrs.h
new file mode 100644
index 000000000..70dfefe8f
--- /dev/null
+++ b/boehm-gc/include/private/gc_hdrs.h
@@ -0,0 +1,233 @@
+/* 
+ * Copyright 1988, 1989 Hans-J. Boehm, Alan J. Demers
+ * Copyright (c) 1991-1994 by Xerox Corporation.  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.
+ */
+/* Boehm, July 11, 1995 11:54 am PDT */
+# ifndef GC_HEADERS_H
+# define GC_HEADERS_H
+typedef struct hblkhdr hdr;
+
+# if CPP_WORDSZ != 32 && CPP_WORDSZ < 36
+	--> Get a real machine.
+# endif
+
+/*
+ * The 2 level tree data structure that is used to find block headers.
+ * If there are more than 32 bits in a pointer, the top level is a hash
+ * table.
+ *
+ * This defines HDR, GET_HDR, and SET_HDR, the main macros used to
+ * retrieve and set object headers.
+ *
+ * Since 5.0 alpha 5, we can also take advantage of a header lookup
+ * cache.  This is a locally declared direct mapped cache, used inside
+ * the marker.  The HC_GET_HDR macro uses and maintains this
+ * cache.  Assuming we get reasonable hit rates, this shaves a few
+ * memory references from each pointer validation.
+ */
+
+# if CPP_WORDSZ > 32
+#   define HASH_TL
+# endif
+
+/* Define appropriate out-degrees for each of the two tree levels	*/
+# ifdef SMALL_CONFIG
+#   define LOG_BOTTOM_SZ 11
+	/* Keep top index size reasonable with smaller blocks. */
+# else
+#   define LOG_BOTTOM_SZ 10
+# endif
+# ifndef HASH_TL
+#   define LOG_TOP_SZ (WORDSZ - LOG_BOTTOM_SZ - LOG_HBLKSIZE)
+# else
+#   define LOG_TOP_SZ 11
+# endif
+# define TOP_SZ (1 << LOG_TOP_SZ)
+# define BOTTOM_SZ (1 << LOG_BOTTOM_SZ)
+
+#ifndef SMALL_CONFIG
+# define USE_HDR_CACHE
+#endif
+
+/* #define COUNT_HDR_CACHE_HITS  */
+
+extern hdr * GC_invalid_header; /* header for an imaginary block 	*/
+				/* containing no objects.		*/
+
+
+/* Check whether p and corresponding hhdr point to long or invalid	*/
+/* object.  If so, advance hhdr	to					*/
+/* beginning of	block, or set hhdr to GC_invalid_header.		*/
+#define ADVANCE(p, hhdr, source) \
+	    { \
+	      hdr * new_hdr = GC_invalid_header; \
+              p = GC_find_start(p, hhdr, &new_hdr); \
+	      hhdr = new_hdr; \
+    	    }
+
+#ifdef USE_HDR_CACHE
+
+# ifdef COUNT_HDR_CACHE_HITS
+    extern word GC_hdr_cache_hits;
+    extern word GC_hdr_cache_misses;
+#   define HC_HIT() ++GC_hdr_cache_hits
+#   define HC_MISS() ++GC_hdr_cache_misses
+# else
+#   define HC_HIT()
+#   define HC_MISS()
+# endif
+
+  typedef struct hce {
+    word block_addr;  	/* right shifted by LOG_HBLKSIZE */
+    hdr * hce_hdr;
+  } hdr_cache_entry;
+
+# define HDR_CACHE_SIZE 8  /* power of 2 */
+
+# define DECLARE_HDR_CACHE \
+	hdr_cache_entry hdr_cache[HDR_CACHE_SIZE]
+
+# define INIT_HDR_CACHE BZERO(hdr_cache, sizeof(hdr_cache));
+
+# define HCE(h) hdr_cache + (((word)(h) >> LOG_HBLKSIZE) & (HDR_CACHE_SIZE-1))
+
+# define HCE_VALID_FOR(hce,h) ((hce) -> block_addr == \
+				((word)(h) >> LOG_HBLKSIZE))
+
+# define HCE_HDR(h) ((hce) -> hce_hdr)
+
+
+/* Analogous to GET_HDR, except that in the case of large objects, it	*/
+/* Returns the header for the object beginning, and updates p.		*/
+/* Returns GC_invalid_header instead of 0.  All of this saves a branch	*/
+/* in the fast path.							*/
+# define HC_GET_HDR(p, hhdr, source) \
+	{ \
+	  hdr_cache_entry * hce = HCE(p); \
+	  if (HCE_VALID_FOR(hce, p)) { \
+	    HC_HIT(); \
+	    hhdr = hce -> hce_hdr; \
+	  } else { \
+	    HC_MISS(); \
+	    GET_HDR(p, hhdr); \
+            if (IS_FORWARDING_ADDR_OR_NIL(hhdr)) { \
+	      ADVANCE(p, hhdr, source); \
+	    } else { \
+	      hce -> block_addr = (word)(p) >> LOG_HBLKSIZE; \
+	      hce -> hce_hdr = hhdr; \
+	    } \
+	  } \
+	}
+
+#else /* !USE_HDR_CACHE */
+
+# define DECLARE_HDR_CACHE
+
+# define INIT_HDR_CACHE
+
+# define HC_GET_HDR(p, hhdr, source) \
+	{ \
+	  GET_HDR(p, hhdr); \
+          if (IS_FORWARDING_ADDR_OR_NIL(hhdr)) { \
+	    ADVANCE(p, hhdr, source); \
+	  } \
+	}
+#endif
+
+typedef struct bi {
+    hdr * index[BOTTOM_SZ];
+	/*
+ 	 * The bottom level index contains one of three kinds of values:
+	 * 0 means we're not responsible for this block,
+	 *   or this is a block other than the first one in a free block.
+	 * 1 < (long)X <= MAX_JUMP means the block starts at least
+	 *        X * HBLKSIZE bytes before the current address.
+	 * A valid pointer points to a hdr structure. (The above can't be
+	 * valid pointers due to the GET_MEM return convention.)
+	 */
+    struct bi * asc_link;	/* All indices are linked in	*/
+    				/* ascending order...		*/
+    struct bi * desc_link;	/* ... and in descending order.	*/
+    word key;			/* high order address bits.	*/
+# ifdef HASH_TL
+    struct bi * hash_link;	/* Hash chain link.		*/
+# endif
+} bottom_index;
+
+/* extern bottom_index GC_all_nils; - really part of GC_arrays */
+
+/* extern bottom_index * GC_top_index []; - really part of GC_arrays */
+				/* Each entry points to a bottom_index.	*/
+				/* On a 32 bit machine, it points to 	*/
+				/* the index for a set of high order	*/
+				/* bits equal to the index.  For longer	*/
+				/* addresses, we hash the high order	*/
+				/* bits to compute the index in 	*/
+				/* GC_top_index, and each entry points	*/
+				/* to a hash chain.			*/
+				/* The last entry in each chain is	*/
+				/* GC_all_nils.				*/
+
+
+# define MAX_JUMP (HBLKSIZE - 1)
+
+# define HDR_FROM_BI(bi, p) \
+		((bi)->index[((word)(p) >> LOG_HBLKSIZE) & (BOTTOM_SZ - 1)])
+# ifndef HASH_TL
+#   define BI(p) (GC_top_index \
+		[(word)(p) >> (LOG_BOTTOM_SZ + LOG_HBLKSIZE)])
+#   define HDR_INNER(p) HDR_FROM_BI(BI(p),p)
+#   ifdef SMALL_CONFIG
+#	define HDR(p) GC_find_header((ptr_t)(p))
+#   else
+#	define HDR(p) HDR_INNER(p)
+#   endif
+#   define GET_BI(p, bottom_indx) (bottom_indx) = BI(p)
+#   define GET_HDR(p, hhdr) (hhdr) = HDR(p)
+#   define SET_HDR(p, hhdr) HDR_INNER(p) = (hhdr)
+#   define GET_HDR_ADDR(p, ha) (ha) = &(HDR_INNER(p))
+# else /* hash */
+/*  Hash function for tree top level */
+#   define TL_HASH(hi) ((hi) & (TOP_SZ - 1))
+/*  Set bottom_indx to point to the bottom index for address p */
+#   define GET_BI(p, bottom_indx) \
+	{ \
+	    register word hi = \
+	        (word)(p) >> (LOG_BOTTOM_SZ + LOG_HBLKSIZE); \
+	    register bottom_index * _bi = GC_top_index[TL_HASH(hi)]; \
+	    \
+	    while (_bi -> key != hi && _bi != GC_all_nils) \
+	    	_bi = _bi -> hash_link; \
+	    (bottom_indx) = _bi; \
+	}
+#   define GET_HDR_ADDR(p, ha) \
+	{ \
+	    register bottom_index * bi; \
+	    \
+	    GET_BI(p, bi);	\
+	    (ha) = &(HDR_FROM_BI(bi, p)); \
+	}
+#   define GET_HDR(p, hhdr) { register hdr ** _ha; GET_HDR_ADDR(p, _ha); \
+			      (hhdr) = *_ha; }
+#   define SET_HDR(p, hhdr) { register hdr ** _ha; GET_HDR_ADDR(p, _ha); \
+			      *_ha = (hhdr); }
+#   define HDR(p) GC_find_header((ptr_t)(p))
+# endif
+			    
+/* Is the result a forwarding address to someplace closer to the	*/
+/* beginning of the block or NIL?					*/
+# define IS_FORWARDING_ADDR_OR_NIL(hhdr) ((unsigned long) (hhdr) <= MAX_JUMP)
+
+/* Get an HBLKSIZE aligned address closer to the beginning of the block */
+/* h.  Assumes hhdr == HDR(h) and IS_FORWARDING_ADDR(hhdr).		*/
+# define FORWARDED_ADDR(h, hhdr) ((struct hblk *)(h) - (unsigned long)(hhdr))
+# endif /*  GC_HEADERS_H */
diff --git a/boehm-gc/include/private/gc_locks.h b/boehm-gc/include/private/gc_locks.h
new file mode 100644
index 000000000..54cfdb92d
--- /dev/null
+++ b/boehm-gc/include/private/gc_locks.h
@@ -0,0 +1,684 @@
+/* 
+ * Copyright 1988, 1989 Hans-J. Boehm, Alan J. Demers
+ * Copyright (c) 1991-1994 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.
+ */
+
+#ifndef GC_LOCKS_H
+#define GC_LOCKS_H
+
+/*
+ * Mutual exclusion between allocator/collector routines.
+ * Needed if there is more than one allocator thread.
+ * FASTLOCK() is assumed to try to acquire the lock in a cheap and
+ * dirty way that is acceptable for a few instructions, e.g. by
+ * inhibiting preemption.  This is assumed to have succeeded only
+ * if a subsequent call to FASTLOCK_SUCCEEDED() returns TRUE.
+ * FASTUNLOCK() is called whether or not FASTLOCK_SUCCEEDED().
+ * If signals cannot be tolerated with the FASTLOCK held, then
+ * FASTLOCK should disable signals.  The code executed under
+ * FASTLOCK is otherwise immune to interruption, provided it is
+ * not restarted.
+ * DCL_LOCK_STATE declares any local variables needed by LOCK and UNLOCK
+ * and/or DISABLE_SIGNALS and ENABLE_SIGNALS and/or FASTLOCK.
+ * (There is currently no equivalent for FASTLOCK.)
+ *
+ * In the PARALLEL_MARK case, we also need to define a number of
+ * other inline finctions here:
+ *   GC_bool GC_compare_and_exchange( volatile GC_word *addr,
+ *   				      GC_word old, GC_word new )
+ *   GC_word GC_atomic_add( volatile GC_word *addr, GC_word how_much )
+ *   void GC_memory_barrier( )
+ *   
+ */  
+# ifdef THREADS
+   void GC_noop1 GC_PROTO((word));
+#  ifdef PCR_OBSOLETE	/* Faster, but broken with multiple lwp's	*/
+#    include  "th/PCR_Th.h"
+#    include  "th/PCR_ThCrSec.h"
+     extern struct PCR_Th_MLRep GC_allocate_ml;
+#    define DCL_LOCK_STATE  PCR_sigset_t GC_old_sig_mask
+#    define LOCK() PCR_Th_ML_Acquire(&GC_allocate_ml) 
+#    define UNLOCK() PCR_Th_ML_Release(&GC_allocate_ml)
+#    define UNLOCK() PCR_Th_ML_Release(&GC_allocate_ml)
+#    define FASTLOCK() PCR_ThCrSec_EnterSys()
+     /* Here we cheat (a lot): */
+#        define FASTLOCK_SUCCEEDED() (*(int *)(&GC_allocate_ml) == 0)
+		/* TRUE if nobody currently holds the lock */
+#    define FASTUNLOCK() PCR_ThCrSec_ExitSys()
+#  endif
+#  ifdef PCR
+#    include <base/PCR_Base.h>
+#    include <th/PCR_Th.h>
+     extern PCR_Th_ML GC_allocate_ml;
+#    define DCL_LOCK_STATE \
+	 PCR_ERes GC_fastLockRes; PCR_sigset_t GC_old_sig_mask
+#    define LOCK() PCR_Th_ML_Acquire(&GC_allocate_ml)
+#    define UNLOCK() PCR_Th_ML_Release(&GC_allocate_ml)
+#    define FASTLOCK() (GC_fastLockRes = PCR_Th_ML_Try(&GC_allocate_ml))
+#    define FASTLOCK_SUCCEEDED() (GC_fastLockRes == PCR_ERes_okay)
+#    define FASTUNLOCK()  {\
+        if( FASTLOCK_SUCCEEDED() ) PCR_Th_ML_Release(&GC_allocate_ml); }
+#  endif
+#  ifdef SRC_M3
+     extern GC_word RT0u__inCritical;
+#    define LOCK() RT0u__inCritical++
+#    define UNLOCK() RT0u__inCritical--
+#  endif
+#  ifdef GC_SOLARIS_THREADS
+#    include <thread.h>
+#    include <signal.h>
+     extern mutex_t GC_allocate_ml;
+#    define LOCK() mutex_lock(&GC_allocate_ml);
+#    define UNLOCK() mutex_unlock(&GC_allocate_ml);
+#  endif
+
+/* Try to define GC_TEST_AND_SET and a matching GC_CLEAR for spin lock	*/
+/* acquisition and release.  We need this for correct operation of the	*/
+/* incremental GC.							*/
+#  ifdef __GNUC__
+#    if defined(I386)
+       inline static int GC_test_and_set(volatile unsigned int *addr) {
+	  int oldval;
+	  /* Note: the "xchg" instruction does not need a "lock" prefix */
+	  __asm__ __volatile__("xchgl %0, %1"
+		: "=r"(oldval), "=m"(*(addr))
+		: "0"(1), "m"(*(addr)) : "memory");
+	  return oldval;
+       }
+#      define GC_TEST_AND_SET_DEFINED
+#    endif
+#    if defined(IA64)
+#     include <ia64intrin.h>
+       inline static int GC_test_and_set(volatile unsigned int *addr) {
+	  return __sync_lock_test_and_set(addr, 1);
+       }
+#      define GC_TEST_AND_SET_DEFINED
+       inline static void GC_clear(volatile unsigned int *addr) {
+	  *addr = 0;
+       }
+#      define GC_CLEAR_DEFINED
+#    endif
+#    ifdef SPARC
+       inline static int GC_test_and_set(volatile unsigned int *addr) {
+	 int oldval;
+
+	 __asm__ __volatile__("ldstub %1,%0"
+	 : "=r"(oldval), "=m"(*addr)
+	 : "m"(*addr) : "memory");
+	 return oldval;
+       }
+#      define GC_TEST_AND_SET_DEFINED
+#    endif
+#    ifdef M68K
+       /* Contributed by Tony Mantler.  I'm not sure how well it was	*/
+       /* tested.							*/
+       inline static int GC_test_and_set(volatile unsigned int *addr) {
+          char oldval; /* this must be no longer than 8 bits */
+
+          /* The return value is semi-phony. */
+          /* 'tas' sets bit 7 while the return */
+          /* value pretends bit 0 was set */
+          __asm__ __volatile__(
+                 "tas %1@; sne %0; negb %0"
+                 : "=d" (oldval)
+                 : "a" (addr) : "memory");
+          return oldval;
+       }
+#      define GC_TEST_AND_SET_DEFINED
+#    endif
+#    if defined(POWERPC)
+#     define GC_TEST_AND_SET_DEFINED
+#     define GC_CLEAR_DEFINED
+#     if (__GNUC__>4)||((__GNUC__==4)&&(__GNUC_MINOR__>=4))
+#       define GC_test_and_set(addr) __sync_lock_test_and_set (addr, 1)
+#       define GC_clear(addr) __sync_lock_release (addr)
+#     else
+        inline static int GC_test_and_set(volatile unsigned int *addr) {
+          int oldval;
+          int temp = 1; /* locked value */
+
+          __asm__ __volatile__(
+               "\n1:\n"
+	       "\tlwarx %0,%y3\n"      /* load and reserve, 32-bits      */
+               "\tcmpwi %0, 0\n"       /* if load is                     */
+               "\tbne 2f\n"            /*   non-zero, return already set */
+               "\tstwcx. %2,%y3\n"     /* else store conditional         */
+               "\tbne- 1b\n"           /* retry if lost reservation      */
+               "\tsync\n"              /* import barrier                 */
+               "2:\t\n"                /* oldval is zero if we set       */
+              : "=&r"(oldval), "=m"(addr)
+              : "r"(temp), "Z"(addr)
+              : "cr0","memory");
+          return oldval;
+        }
+      inline static void GC_clear(volatile unsigned int *addr) {
+	__asm__ __volatile__("lwsync" : : : "memory");
+        *(addr) = 0;
+      }
+#    endif
+#    endif
+#    if defined(ALPHA) 
+        inline static int GC_test_and_set(volatile unsigned int * addr)
+        {
+          unsigned long oldvalue;
+          unsigned long temp;
+
+          __asm__ __volatile__(
+                             "1:     ldl_l %0,%1\n"
+                             "       and %0,%3,%2\n"
+                             "       bne %2,2f\n"
+                             "       xor %0,%3,%0\n"
+                             "       stl_c %0,%1\n"
+#	ifdef __ELF__
+                             "       beq %0,3f\n"
+#	else
+                             "       beq %0,1b\n"
+#	endif
+                             "       mb\n"
+                             "2:\n"
+#	ifdef __ELF__
+                             ".section .text2,\"ax\"\n"
+                             "3:     br 1b\n"
+                             ".previous"
+#	endif
+                             :"=&r" (temp), "=m" (*addr), "=&r" (oldvalue)
+                             :"Ir" (1), "m" (*addr)
+			     :"memory");
+
+          return oldvalue;
+        }
+#       define GC_TEST_AND_SET_DEFINED
+        inline static void GC_clear(volatile unsigned int *addr) {
+          __asm__ __volatile__("mb" : : : "memory");
+          *(addr) = 0;
+        }
+#       define GC_CLEAR_DEFINED
+#    endif /* ALPHA */
+#    ifdef ARM32
+#     define GC_TEST_AND_SET_DEFINED
+#     if (__GNUC__>4)||((__GNUC__==4)&&(__GNUC_MINOR__>=5)) && defined(__ARM_EABI__)
+#       define GC_CLEAR_DEFINED
+#       define GC_test_and_set(addr) __sync_lock_test_and_set (addr, 1)
+#       define GC_clear(addr) __sync_lock_release (addr)
+#     else
+        inline static int GC_test_and_set(volatile unsigned int *addr) {
+          int oldval;
+          /* SWP on ARM is very similar to XCHG on x86.  Doesn't lock the
+           * bus because there are no SMP ARM machines.  If/when there are,
+           * this code will likely need to be updated. */
+          /* See linuxthreads/sysdeps/arm/pt-machine.h in glibc-2.1 */
+          __asm__ __volatile__("swp %0, %1, [%2]"
+      		  	     : "=r"(oldval)
+      			     : "0"(1), "r"(addr)
+			     : "memory");
+          return oldval;
+        }
+#     endif
+#    endif /* ARM32 */
+#    ifdef CRIS
+        inline static int GC_test_and_set(volatile unsigned int *addr) {
+	  /* Ripped from linuxthreads/sysdeps/cris/pt-machine.h.	*/
+	  /* Included with Hans-Peter Nilsson's permission.		*/
+	  register unsigned long int ret;
+
+	  /* Note the use of a dummy output of *addr to expose the write.
+	   * The memory barrier is to stop *other* writes being moved past
+	   * this code.
+	   */
+	    __asm__ __volatile__("clearf\n"
+        		         "0:\n\t"
+                    		 "movu.b [%2],%0\n\t"
+                    		 "ax\n\t"
+                    		 "move.b %3,[%2]\n\t"
+                    		 "bwf 0b\n\t"
+                    		 "clearf"
+                    		 : "=&r" (ret), "=m" (*addr)
+                    		 : "r" (addr), "r" ((int) 1), "m" (*addr)
+                    		 : "memory");
+	    return ret;
+        }
+#       define GC_TEST_AND_SET_DEFINED
+#    endif /* CRIS */
+#    ifdef S390
+       inline static int GC_test_and_set(volatile unsigned int *addr) {
+         int ret;
+         __asm__ __volatile__ (
+          "     l     %0,0(%2)\n"
+          "0:   cs    %0,%1,0(%2)\n"
+          "     jl    0b"
+          : "=&d" (ret)
+          : "d" (1), "a" (addr)
+          : "cc", "memory");
+         return ret;
+       }
+#    endif
+#  endif /* __GNUC__ */
+#  if (defined(ALPHA) && !defined(__GNUC__))
+#    ifndef OSF1
+	--> We currently assume that if gcc is not used, we are
+	--> running under Tru64.
+#    endif
+#    include <machine/builtins.h>
+#    include <c_asm.h>
+#    define GC_test_and_set(addr) __ATOMIC_EXCH_LONG(addr, 1)
+#    define GC_TEST_AND_SET_DEFINED
+#    define GC_clear(addr) { asm("mb"); *(volatile unsigned *)addr = 0; }
+#    define GC_CLEAR_DEFINED
+#  endif
+#  if defined(MSWIN32)
+#    define GC_test_and_set(addr) InterlockedExchange((LPLONG)addr,1)
+#    define GC_TEST_AND_SET_DEFINED
+#  endif
+#  ifdef MIPS
+#    ifdef LINUX
+#      include <sys/tas.h>
+#      define GC_test_and_set(addr) _test_and_set((int *) addr,1)
+#      define GC_TEST_AND_SET_DEFINED
+#    elif __mips < 3 || !(defined (_ABIN32) || defined(_ABI64)) \
+	|| !defined(_COMPILER_VERSION) || _COMPILER_VERSION < 700
+#	 ifdef __GNUC__
+#          define GC_test_and_set(addr) _test_and_set((void *)addr,1)
+#	 else
+#          define GC_test_and_set(addr) test_and_set((void *)addr,1)
+#	 endif
+#    else
+#	 include <sgidefs.h>
+#	 include <mutex.h>
+#	 define GC_test_and_set(addr) __test_and_set32((void *)addr,1)
+#	 define GC_clear(addr) __lock_release(addr);
+#	 define GC_CLEAR_DEFINED
+#    endif
+#    define GC_TEST_AND_SET_DEFINED
+#  endif /* MIPS */
+#  if defined(_AIX)
+#    include <sys/atomic_op.h>
+#    if (defined(_POWER) || defined(_POWERPC)) 
+#      if defined(__GNUC__)  
+         inline static void GC_memsync() {
+           __asm__ __volatile__ ("sync" : : : "memory");
+         }
+#      else
+#        ifndef inline
+#          define inline __inline
+#        endif
+#        pragma mc_func GC_memsync { \
+           "7c0004ac" /* sync (same opcode used for dcs)*/ \
+         }
+#      endif
+#    else 
+#    error dont know how to memsync
+#    endif
+     inline static int GC_test_and_set(volatile unsigned int * addr) {
+          int oldvalue = 0;
+          if (compare_and_swap((void *)addr, &oldvalue, 1)) {
+            GC_memsync();
+            return 0;
+          } else return 1;
+     }
+#    define GC_TEST_AND_SET_DEFINED
+     inline static void GC_clear(volatile unsigned int *addr) {
+          GC_memsync();
+          *(addr) = 0;
+     }
+#    define GC_CLEAR_DEFINED
+
+#  endif
+#  if 0 /* defined(HP_PA) */
+     /* The official recommendation seems to be to not use ldcw from	*/
+     /* user mode.  Since multithreaded incremental collection doesn't	*/
+     /* work anyway on HP_PA, this shouldn't be a major loss.		*/
+
+     /* "set" means 0 and "clear" means 1 here.		*/
+#    define GC_test_and_set(addr) !GC_test_and_clear(addr);
+#    define GC_TEST_AND_SET_DEFINED
+#    define GC_clear(addr) GC_noop1((word)(addr)); *(volatile unsigned int *)addr = 1;
+	/* The above needs a memory barrier! */
+#    define GC_CLEAR_DEFINED
+#  endif
+#  if defined(GC_TEST_AND_SET_DEFINED) && !defined(GC_CLEAR_DEFINED)
+#    ifdef __GNUC__
+       inline static void GC_clear(volatile unsigned int *addr) {
+         /* Try to discourage gcc from moving anything past this. */
+         __asm__ __volatile__(" " : : : "memory");
+         *(addr) = 0;
+       }
+#    else
+	    /* The function call in the following should prevent the	*/
+	    /* compiler from moving assignments to below the UNLOCK.	*/
+#      define GC_clear(addr) GC_noop1((word)(addr)); \
+			     *((volatile unsigned int *)(addr)) = 0;
+#    endif
+#    define GC_CLEAR_DEFINED
+#  endif /* !GC_CLEAR_DEFINED */
+
+#  if !defined(GC_TEST_AND_SET_DEFINED)
+#    define USE_PTHREAD_LOCKS
+#  endif
+
+#  if defined(GC_PTHREADS) && !defined(GC_SOLARIS_THREADS) \
+      && !defined(GC_WIN32_THREADS)
+#    define NO_THREAD (pthread_t)(-1)
+#    include <pthread.h>
+#    if defined(PARALLEL_MARK) 
+      /* We need compare-and-swap to update mark bits, where it's	*/
+      /* performance critical.  If USE_MARK_BYTES is defined, it is	*/
+      /* no longer needed for this purpose.  However we use it in	*/
+      /* either case to implement atomic fetch-and-add, though that's	*/
+      /* less performance critical, and could perhaps be done with	*/
+      /* a lock.							*/
+#     if defined(GENERIC_COMPARE_AND_SWAP)
+	/* Probably not useful, except for debugging.	*/
+	/* We do use GENERIC_COMPARE_AND_SWAP on PA_RISC, but we 	*/
+	/* minimize its use.						*/
+	extern pthread_mutex_t GC_compare_and_swap_lock;
+
+	/* Note that if GC_word updates are not atomic, a concurrent 	*/
+	/* reader should acquire GC_compare_and_swap_lock.  On 		*/
+	/* currently supported platforms, such updates are atomic.	*/
+	extern GC_bool GC_compare_and_exchange(volatile GC_word *addr,
+					       GC_word old, GC_word new_val);
+#     endif /* GENERIC_COMPARE_AND_SWAP */
+#     if defined(I386)
+#      if !defined(GENERIC_COMPARE_AND_SWAP)
+         /* Returns TRUE if the comparison succeeded. */
+         inline static GC_bool GC_compare_and_exchange(volatile GC_word *addr,
+		  				       GC_word old,
+						       GC_word new_val) 
+         {
+	   char result;
+	   __asm__ __volatile__("lock; cmpxchgl %2, %0; setz %1"
+	    	: "+m"(*(addr)), "=r"(result)
+		: "r" (new_val), "a"(old) : "memory");
+	   return (GC_bool) result;
+         }
+#      endif /* !GENERIC_COMPARE_AND_SWAP */
+       inline static void GC_memory_barrier()
+       {
+	 /* We believe the processor ensures at least processor	*/
+	 /* consistent ordering.  Thus a compiler barrier	*/
+	 /* should suffice.					*/
+         __asm__ __volatile__("" : : : "memory");
+       }
+#     endif /* I386 */
+
+#     if defined(POWERPC)
+#      if !defined(GENERIC_COMPARE_AND_SWAP)
+#       if CPP_WORDSZ == 64
+        /* Returns TRUE if the comparison succeeded. */
+        inline static GC_bool GC_compare_and_exchange(volatile GC_word *addr,
+            GC_word old, GC_word new_val) 
+        {
+            unsigned long result, dummy;
+            __asm__ __volatile__(
+                "1:\tldarx %0,0,%5\n"
+                  "\tcmpd %0,%4\n"
+                  "\tbne  2f\n"
+                  "\tstdcx. %3,0,%2\n"
+                  "\tbne- 1b\n"
+                  "\tsync\n"
+                  "\tli %1, 1\n"
+                  "\tb 3f\n"
+                "2:\tli %1, 0\n"
+                "3:\t\n"
+                :  "=&r" (dummy), "=r" (result), "=p" (addr)
+                :  "r" (new_val), "r" (old), "2"(addr)
+                : "cr0","memory");
+            return (GC_bool) result;
+        }
+#       else
+        /* Returns TRUE if the comparison succeeded. */
+        inline static GC_bool GC_compare_and_exchange(volatile GC_word *addr,
+            GC_word old, GC_word new_val) 
+        {
+            int result, dummy;
+            __asm__ __volatile__(
+                "1:\tlwarx %0,0,%5\n"
+                  "\tcmpw %0,%4\n"
+                  "\tbne  2f\n"
+                  "\tstwcx. %3,0,%2\n"
+                  "\tbne- 1b\n"
+                  "\tsync\n"
+                  "\tli %1, 1\n"
+                  "\tb 3f\n"
+                "2:\tli %1, 0\n"
+                "3:\t\n"
+                :  "=&r" (dummy), "=r" (result), "=p" (addr)
+                :  "r" (new_val), "r" (old), "2"(addr)
+                : "cr0","memory");
+            return (GC_bool) result;
+        }
+#       endif
+#      endif /* !GENERIC_COMPARE_AND_SWAP */
+        inline static void GC_memory_barrier()
+        {
+            __asm__ __volatile__("sync" : : : "memory");
+        }
+#     endif /* POWERPC */
+
+#     if defined(IA64)
+#      if !defined(GENERIC_COMPARE_AND_SWAP)
+         inline static GC_bool GC_compare_and_exchange(volatile GC_word *addr,
+						       GC_word old,
+						       GC_word new_val) 
+	 {
+	   return __sync_bool_compare_and_swap (addr, old, new_val);
+         }
+#      endif /* !GENERIC_COMPARE_AND_SWAP */
+#      if 0
+	/* Shouldn't be needed; we use volatile stores instead. */
+        inline static void GC_memory_barrier()
+        {
+          __sync_synchronize ();
+        }
+#      endif /* 0 */
+#     endif /* IA64 */
+#     if defined(ALPHA)
+#      if !defined(GENERIC_COMPARE_AND_SWAP)
+#        if defined(__GNUC__)
+           inline static GC_bool GC_compare_and_exchange(volatile GC_word *addr,
+						         GC_word old, GC_word new_val) 
+	   {
+	     unsigned long was_equal;
+             unsigned long temp;
+
+             __asm__ __volatile__(
+                             "1:     ldq_l %0,%1\n"
+                             "       cmpeq %0,%4,%2\n"
+			     "	     mov %3,%0\n"
+                             "       beq %2,2f\n"
+                             "       stq_c %0,%1\n"
+                             "       beq %0,1b\n"
+                             "2:\n"
+                             "       mb\n"
+                             :"=&r" (temp), "=m" (*addr), "=&r" (was_equal)
+                             : "r" (new_val), "Ir" (old)
+			     :"memory");
+             return was_equal;
+           }
+#        else /* !__GNUC__ */
+           inline static GC_bool GC_compare_and_exchange(volatile GC_word *addr,
+						         GC_word old, GC_word new_val) 
+	  {
+	    return __CMP_STORE_QUAD(addr, old, new_val, addr);
+          }
+#        endif /* !__GNUC__ */
+#      endif /* !GENERIC_COMPARE_AND_SWAP */
+#      ifdef __GNUC__
+         inline static void GC_memory_barrier()
+         {
+           __asm__ __volatile__("mb" : : : "memory");
+         }
+#      else
+#	 define GC_memory_barrier() asm("mb")
+#      endif /* !__GNUC__ */
+#     endif /* ALPHA */
+#     if defined(S390)
+#      if !defined(GENERIC_COMPARE_AND_SWAP)
+         inline static GC_bool GC_compare_and_exchange(volatile C_word *addr,
+                                         GC_word old, GC_word new_val)
+         {
+           int retval;
+           __asm__ __volatile__ (
+#            ifndef __s390x__
+               "     cs  %1,%2,0(%3)\n"
+#            else
+               "     csg %1,%2,0(%3)\n"
+#            endif
+             "     ipm %0\n"
+             "     srl %0,28\n"
+             : "=&d" (retval), "+d" (old)
+             : "d" (new_val), "a" (addr)
+             : "cc", "memory");
+           return retval == 0;
+         }
+#      endif
+#     endif
+#     if !defined(GENERIC_COMPARE_AND_SWAP)
+        /* Returns the original value of *addr.	*/
+        inline static GC_word GC_atomic_add(volatile GC_word *addr,
+					    GC_word how_much)
+        {
+	  GC_word old;
+	  do {
+	    old = *addr;
+	  } while (!GC_compare_and_exchange(addr, old, old+how_much));
+          return old;
+        }
+#     else /* GENERIC_COMPARE_AND_SWAP */
+	/* So long as a GC_word can be atomically updated, it should	*/
+	/* be OK to read *addr without a lock.				*/
+	extern GC_word GC_atomic_add(volatile GC_word *addr, GC_word how_much);
+#     endif /* GENERIC_COMPARE_AND_SWAP */
+
+#    endif /* PARALLEL_MARK */
+
+#    if !defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_LOCKS)
+      /* In the THREAD_LOCAL_ALLOC case, the allocation lock tends to	*/
+      /* be held for long periods, if it is held at all.  Thus spinning	*/
+      /* and sleeping for fixed periods are likely to result in 	*/
+      /* significant wasted time.  We thus rely mostly on queued locks. */
+#     define USE_SPIN_LOCK
+      extern volatile unsigned int GC_allocate_lock;
+      extern void GC_lock(void);
+	/* Allocation lock holder.  Only set if acquired by client through */
+	/* GC_call_with_alloc_lock.					   */
+#     ifdef GC_ASSERTIONS
+#        define LOCK() \
+		{ if (GC_test_and_set(&GC_allocate_lock)) GC_lock(); \
+		  SET_LOCK_HOLDER(); }
+#        define UNLOCK() \
+		{ GC_ASSERT(I_HOLD_LOCK()); UNSET_LOCK_HOLDER(); \
+	          GC_clear(&GC_allocate_lock); }
+#     else
+#        define LOCK() \
+		{ if (GC_test_and_set(&GC_allocate_lock)) GC_lock(); }
+#        define UNLOCK() \
+		GC_clear(&GC_allocate_lock)
+#     endif /* !GC_ASSERTIONS */
+#     if 0
+	/* Another alternative for OSF1 might be:		*/
+#       include <sys/mman.h>
+        extern msemaphore GC_allocate_semaphore;
+#       define LOCK() { if (msem_lock(&GC_allocate_semaphore, MSEM_IF_NOWAIT) \
+ 			    != 0) GC_lock(); else GC_allocate_lock = 1; }
+        /* The following is INCORRECT, since the memory model is too weak. */
+	/* Is this true?  Presumably msem_unlock has the right semantics?  */
+	/*		- HB						   */
+#       define UNLOCK() { GC_allocate_lock = 0; \
+                          msem_unlock(&GC_allocate_semaphore, 0); }
+#     endif /* 0 */
+#    else /* THREAD_LOCAL_ALLOC  || USE_PTHREAD_LOCKS */
+#      ifndef USE_PTHREAD_LOCKS
+#        define USE_PTHREAD_LOCKS
+#      endif
+#    endif /* THREAD_LOCAL_ALLOC */
+#   ifdef USE_PTHREAD_LOCKS
+#      include <pthread.h>
+       extern pthread_mutex_t GC_allocate_ml;
+#      ifdef GC_ASSERTIONS
+#        define LOCK() \
+		{ GC_lock(); \
+		  SET_LOCK_HOLDER(); }
+#        define UNLOCK() \
+		{ GC_ASSERT(I_HOLD_LOCK()); UNSET_LOCK_HOLDER(); \
+	          pthread_mutex_unlock(&GC_allocate_ml); }
+#      else /* !GC_ASSERTIONS */
+#        if defined(NO_PTHREAD_TRYLOCK)
+#          define LOCK() GC_lock();
+#        else /* !defined(NO_PTHREAD_TRYLOCK) */
+#        define LOCK() \
+	   { if (0 != pthread_mutex_trylock(&GC_allocate_ml)) GC_lock(); }
+#        endif
+#        define UNLOCK() pthread_mutex_unlock(&GC_allocate_ml)
+#      endif /* !GC_ASSERTIONS */
+#   endif /* USE_PTHREAD_LOCKS */
+#   define SET_LOCK_HOLDER() GC_lock_holder = pthread_self()
+#   define UNSET_LOCK_HOLDER() GC_lock_holder = NO_THREAD
+#   define I_HOLD_LOCK() (pthread_equal(GC_lock_holder, pthread_self()))
+    extern VOLATILE GC_bool GC_collecting;
+#   define ENTER_GC() GC_collecting = 1;
+#   define EXIT_GC() GC_collecting = 0;
+    extern void GC_lock(void);
+    extern pthread_t GC_lock_holder;
+#   ifdef GC_ASSERTIONS
+      extern pthread_t GC_mark_lock_holder;
+#   endif
+#  endif /* GC_PTHREADS with linux_threads.c implementation */
+#  if defined(GC_WIN32_THREADS)
+#    if defined(GC_PTHREADS)
+#      include <pthread.h>
+       extern pthread_mutex_t GC_allocate_ml;
+#      define LOCK()   pthread_mutex_lock(&GC_allocate_ml)
+#      define UNLOCK() pthread_mutex_unlock(&GC_allocate_ml)
+#    else
+#      include <windows.h>
+       GC_API CRITICAL_SECTION GC_allocate_ml;
+#      define LOCK() EnterCriticalSection(&GC_allocate_ml);
+#      define UNLOCK() LeaveCriticalSection(&GC_allocate_ml);
+#    endif
+#  endif
+#  ifndef SET_LOCK_HOLDER
+#      define SET_LOCK_HOLDER()
+#      define UNSET_LOCK_HOLDER()
+#      define I_HOLD_LOCK() FALSE
+		/* Used on platforms were locks can be reacquired,	*/
+		/* so it doesn't matter if we lie.			*/
+#  endif
+# else /* !THREADS */
+#    define LOCK()
+#    define UNLOCK()
+# endif /* !THREADS */
+# ifndef SET_LOCK_HOLDER
+#   define SET_LOCK_HOLDER()
+#   define UNSET_LOCK_HOLDER()
+#   define I_HOLD_LOCK() FALSE
+		/* Used on platforms were locks can be reacquired,	*/
+		/* so it doesn't matter if we lie.			*/
+# endif
+# ifndef ENTER_GC
+#   define ENTER_GC()
+#   define EXIT_GC()
+# endif
+
+# ifndef DCL_LOCK_STATE
+#   define DCL_LOCK_STATE
+# endif
+# ifndef FASTLOCK
+#   define FASTLOCK() LOCK()
+#   define FASTLOCK_SUCCEEDED() TRUE
+#   define FASTUNLOCK() UNLOCK()
+# endif
+
+#endif /* GC_LOCKS_H */
diff --git a/boehm-gc/include/private/gc_pmark.h b/boehm-gc/include/private/gc_pmark.h
new file mode 100644
index 000000000..6b15d8ad2
--- /dev/null
+++ b/boehm-gc/include/private/gc_pmark.h
@@ -0,0 +1,394 @@
+/*
+ * Copyright (c) 1991-1994 by Xerox Corporation.  All rights reserved.
+ * Copyright (c) 2001 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.
+ *
+ */
+
+/* Private declarations of GC marker data structures and macros */
+
+/*
+ * Declarations of mark stack.  Needed by marker and client supplied mark
+ * routines.  Transitively include gc_priv.h.
+ * (Note that gc_priv.h should not be included before this, since this
+ * includes dbg_mlc.h, which wants to include gc_priv.h AFTER defining
+ * I_HIDE_POINTERS.)
+ */
+#ifndef GC_PMARK_H
+# define GC_PMARK_H
+
+# include "gc.h"	/* For configuration */
+
+# if defined(KEEP_BACK_PTRS) || defined(PRINT_BLACK_LIST)
+#   include "dbg_mlc.h"
+# endif
+# ifndef GC_MARK_H
+#   include "../gc_mark.h"
+# endif
+# ifndef GC_PRIVATE_H
+#   include "gc_priv.h"
+# endif
+
+/* The real declarations of the following is in gc_priv.h, so that	*/
+/* we can avoid scanning the following table.				*/
+/*
+extern mark_proc GC_mark_procs[MAX_MARK_PROCS];
+*/
+
+/*
+ * Mark descriptor stuff that should remain private for now, mostly
+ * because it's hard to export WORDSZ without including gcconfig.h.
+ */
+# define BITMAP_BITS (WORDSZ - GC_DS_TAG_BITS)
+# define PROC(descr) \
+	(GC_mark_procs[((descr) >> GC_DS_TAG_BITS) & (GC_MAX_MARK_PROCS-1)])
+# define ENV(descr) \
+	((descr) >> (GC_DS_TAG_BITS + GC_LOG_MAX_MARK_PROCS))
+# define MAX_ENV \
+  	(((word)1 << (WORDSZ - GC_DS_TAG_BITS - GC_LOG_MAX_MARK_PROCS)) - 1)
+
+
+extern word GC_n_mark_procs;
+
+/* Number of mark stack entries to discard on overflow.	*/
+#define GC_MARK_STACK_DISCARDS (INITIAL_MARK_STACK_SIZE/8)
+
+typedef struct GC_ms_entry {
+    GC_word * mse_start;   /* First word of object */
+    GC_word mse_descr;	/* Descriptor; low order two bits are tags,	*/
+    			/* identifying the upper 30 bits as one of the	*/
+    			/* following:					*/
+} mse;
+
+extern word GC_mark_stack_size;
+
+extern mse * GC_mark_stack_limit;
+
+#ifdef PARALLEL_MARK
+  extern mse * VOLATILE GC_mark_stack_top;
+#else
+  extern mse * GC_mark_stack_top;
+#endif
+
+extern mse * GC_mark_stack;
+
+#ifdef PARALLEL_MARK
+    /*
+     * Allow multiple threads to participate in the marking process.
+     * This works roughly as follows:
+     *  The main mark stack never shrinks, but it can grow.
+     *
+     *	The initiating threads holds the GC lock, and sets GC_help_wanted.
+     *  
+     *  Other threads:
+     *     1) update helper_count (while holding mark_lock.)
+     *	   2) allocate a local mark stack
+     *     repeatedly:
+     *		3) Steal a global mark stack entry by atomically replacing
+     *		   its descriptor with 0.
+     *		4) Copy it to the local stack.
+     *	        5) Mark on the local stack until it is empty, or
+     *		   it may be profitable to copy it back.
+     *	        6) If necessary, copy local stack to global one,
+     *		   holding mark lock.
+     *    7) Stop when the global mark stack is empty.
+     *    8) decrement helper_count (holding mark_lock).
+     *
+     * This is an experiment to see if we can do something along the lines
+     * of the University of Tokyo SGC in a less intrusive, though probably
+     * also less performant, way.
+     */
+    void GC_do_parallel_mark();
+		/* inititate parallel marking.	*/
+
+    extern GC_bool GC_help_wanted;	/* Protected by mark lock	*/
+    extern unsigned GC_helper_count;	/* Number of running helpers.	*/
+					/* Protected by mark lock	*/
+    extern unsigned GC_active_count;	/* Number of active helpers.	*/
+					/* Protected by mark lock	*/
+					/* May increase and decrease	*/
+					/* within each mark cycle.  But	*/
+					/* once it returns to 0, it	*/
+					/* stays zero for the cycle.	*/
+    /* GC_mark_stack_top is also protected by mark lock.	*/
+    extern mse * VOLATILE GC_first_nonempty;
+					/* Lowest entry on mark stack	*/
+					/* that may be nonempty.	*/
+					/* Updated only by initiating 	*/
+					/* thread.			*/
+    /*
+     * GC_notify_all_marker() is used when GC_help_wanted is first set,
+     * when the last helper becomes inactive,
+     * when something is added to the global mark stack, and just after
+     * GC_mark_no is incremented.
+     * This could be split into multiple CVs (and probably should be to
+     * scale to really large numbers of processors.)
+     */
+#endif /* PARALLEL_MARK */
+
+/* Return a pointer to within 1st page of object.  	*/
+/* Set *new_hdr_p to corr. hdr.				*/
+#ifdef __STDC__
+  ptr_t GC_find_start(ptr_t current, hdr *hhdr, hdr **new_hdr_p);
+#else
+  ptr_t GC_find_start();
+#endif
+
+mse * GC_signal_mark_stack_overflow GC_PROTO((mse *msp));
+
+# ifdef GATHERSTATS
+#   define ADD_TO_ATOMIC(sz) GC_atomic_in_use += (sz)
+#   define ADD_TO_COMPOSITE(sz) GC_composite_in_use += (sz)
+# else
+#   define ADD_TO_ATOMIC(sz)
+#   define ADD_TO_COMPOSITE(sz)
+# endif
+
+/* Push the object obj with corresponding heap block header hhdr onto 	*/
+/* the mark stack.							*/
+# define PUSH_OBJ(obj, hhdr, mark_stack_top, mark_stack_limit) \
+{ \
+    register word _descr = (hhdr) -> hb_descr; \
+        \
+    if (_descr == 0) { \
+    	ADD_TO_ATOMIC((hhdr) -> hb_sz); \
+    } else { \
+        ADD_TO_COMPOSITE((hhdr) -> hb_sz); \
+        mark_stack_top++; \
+        if (mark_stack_top >= mark_stack_limit) { \
+          mark_stack_top = GC_signal_mark_stack_overflow(mark_stack_top); \
+        } \
+        mark_stack_top -> mse_start = (obj); \
+        mark_stack_top -> mse_descr = _descr; \
+    } \
+}
+
+/* Push the contents of current onto the mark stack if it is a valid	*/
+/* ptr to a currently unmarked object.  Mark it.			*/
+/* If we assumed a standard-conforming compiler, we could probably	*/
+/* generate the exit_label transparently.				*/
+# define PUSH_CONTENTS(current, mark_stack_top, mark_stack_limit, \
+		       source, exit_label) \
+{ \
+    hdr * my_hhdr; \
+    ptr_t my_current = current; \
+ \
+    GET_HDR(my_current, my_hhdr); \
+    if (IS_FORWARDING_ADDR_OR_NIL(my_hhdr)) { \
+	 hdr * new_hdr = GC_invalid_header; \
+         my_current = GC_find_start(my_current, my_hhdr, &new_hdr); \
+         my_hhdr = new_hdr; \
+    } \
+    PUSH_CONTENTS_HDR(my_current, mark_stack_top, mark_stack_limit, \
+		  source, exit_label, my_hhdr);	\
+exit_label: ; \
+}
+
+/* As above, but use header cache for header lookup.	*/
+# define HC_PUSH_CONTENTS(current, mark_stack_top, mark_stack_limit, \
+		       source, exit_label) \
+{ \
+    hdr * my_hhdr; \
+    ptr_t my_current = current; \
+ \
+    HC_GET_HDR(my_current, my_hhdr, source); \
+    PUSH_CONTENTS_HDR(my_current, mark_stack_top, mark_stack_limit, \
+		  source, exit_label, my_hhdr);	\
+exit_label: ; \
+}
+
+/* Set mark bit, exit if it was already set.	*/
+
+# ifdef USE_MARK_BYTES
+    /* Unlike the mark bit case, there is a race here, and we may set	*/
+    /* the bit twice in the concurrent case.  This can result in the	*/
+    /* object being pushed twice.  But that's only a performance issue.	*/
+#   define SET_MARK_BIT_EXIT_IF_SET(hhdr,displ,exit_label) \
+    { \
+        register VOLATILE char * mark_byte_addr = \
+				hhdr -> hb_marks + ((displ) >> 1); \
+        register char mark_byte = *mark_byte_addr; \
+          \
+	if (mark_byte) goto exit_label; \
+	*mark_byte_addr = 1;  \
+    } 
+# else
+#   define SET_MARK_BIT_EXIT_IF_SET(hhdr,displ,exit_label) \
+    { \
+        register word * mark_word_addr = hhdr -> hb_marks + divWORDSZ(displ); \
+          \
+        OR_WORD_EXIT_IF_SET(mark_word_addr, (word)1 << modWORDSZ(displ), \
+			    exit_label); \
+    } 
+# endif /* USE_MARK_BYTES */
+
+/* If the mark bit corresponding to current is not set, set it, and 	*/
+/* push the contents of the object on the mark stack.  For a small 	*/
+/* object we assume that current is the (possibly interior) pointer	*/
+/* to the object.  For large objects we assume that current points	*/
+/* to somewhere inside the first page of the object.  If		*/
+/* GC_all_interior_pointers is set, it may have been previously 	*/
+/* adjusted to make that true.						*/
+# define PUSH_CONTENTS_HDR(current, mark_stack_top, mark_stack_limit, \
+		           source, exit_label, hhdr) \
+{ \
+    int displ;  /* Displacement in block; first bytes, then words */ \
+    int map_entry; \
+    \
+    displ = HBLKDISPL(current); \
+    map_entry = MAP_ENTRY((hhdr -> hb_map), displ); \
+    displ = BYTES_TO_WORDS(displ); \
+    if (map_entry > CPP_MAX_OFFSET) { \
+	if (map_entry == OFFSET_TOO_BIG) { \
+	  map_entry = displ % (hhdr -> hb_sz); \
+	  displ -= map_entry; \
+	  if (displ + (hhdr -> hb_sz) > BYTES_TO_WORDS(HBLKSIZE)) { \
+	    GC_ADD_TO_BLACK_LIST_NORMAL((word)current, source); \
+	    goto exit_label; \
+	  } \
+	} else { \
+          GC_ADD_TO_BLACK_LIST_NORMAL((word)current, source); goto exit_label; \
+	} \
+    } else { \
+        displ -= map_entry; \
+    } \
+    GC_ASSERT(displ >= 0 && displ < MARK_BITS_PER_HBLK); \
+    SET_MARK_BIT_EXIT_IF_SET(hhdr, displ, exit_label); \
+    GC_STORE_BACK_PTR((ptr_t)source, (ptr_t)HBLKPTR(current) \
+				      + WORDS_TO_BYTES(displ)); \
+    PUSH_OBJ(((word *)(HBLKPTR(current)) + displ), hhdr, \
+    	     mark_stack_top, mark_stack_limit) \
+}
+
+#if defined(PRINT_BLACK_LIST) || defined(KEEP_BACK_PTRS)
+#   define PUSH_ONE_CHECKED_STACK(p, source) \
+	GC_mark_and_push_stack(p, (ptr_t)(source))
+#else
+#   define PUSH_ONE_CHECKED_STACK(p, source) \
+	GC_mark_and_push_stack(p)
+#endif
+
+/*
+ * Push a single value onto mark stack. Mark from the object pointed to by p.
+ * Invoke FIXUP_POINTER(p) before any further processing.
+ * P is considered valid even if it is an interior pointer.
+ * Previously marked objects are not pushed.  Hence we make progress even
+ * if the mark stack overflows.
+ */
+
+# if NEED_FIXUP_POINTER
+    /* Try both the raw version and the fixed up one.	*/
+#   define GC_PUSH_ONE_STACK(p, source) \
+      if ((ptr_t)(p) >= (ptr_t)GC_least_plausible_heap_addr 	\
+	 && (ptr_t)(p) < (ptr_t)GC_greatest_plausible_heap_addr) {	\
+	 PUSH_ONE_CHECKED_STACK(p, source);	\
+      } \
+      FIXUP_POINTER(p); \
+      if ((ptr_t)(p) >= (ptr_t)GC_least_plausible_heap_addr 	\
+	 && (ptr_t)(p) < (ptr_t)GC_greatest_plausible_heap_addr) {	\
+	 PUSH_ONE_CHECKED_STACK(p, source);	\
+      }
+# else /* !NEED_FIXUP_POINTER */
+#   define GC_PUSH_ONE_STACK(p, source) \
+      if ((ptr_t)(p) >= (ptr_t)GC_least_plausible_heap_addr 	\
+	 && (ptr_t)(p) < (ptr_t)GC_greatest_plausible_heap_addr) {	\
+	 PUSH_ONE_CHECKED_STACK(p, source);	\
+      }
+# endif
+
+
+/*
+ * As above, but interior pointer recognition as for
+ * normal for heap pointers.
+ */
+# define GC_PUSH_ONE_HEAP(p,source) \
+    FIXUP_POINTER(p); \
+    if ((ptr_t)(p) >= (ptr_t)GC_least_plausible_heap_addr 	\
+	 && (ptr_t)(p) < (ptr_t)GC_greatest_plausible_heap_addr) {	\
+	    GC_mark_stack_top = GC_mark_and_push( \
+			    (GC_PTR)(p), GC_mark_stack_top, \
+			    GC_mark_stack_limit, (GC_PTR *)(source)); \
+    }
+
+/* Mark starting at mark stack entry top (incl.) down to	*/
+/* mark stack entry bottom (incl.).  Stop after performing	*/
+/* about one page worth of work.  Return the new mark stack	*/
+/* top entry.							*/
+mse * GC_mark_from GC_PROTO((mse * top, mse * bottom, mse *limit));
+
+#define MARK_FROM_MARK_STACK() \
+	GC_mark_stack_top = GC_mark_from(GC_mark_stack_top, \
+					 GC_mark_stack, \
+					 GC_mark_stack + GC_mark_stack_size);
+
+/*
+ * Mark from one finalizable object using the specified
+ * mark proc. May not mark the object pointed to by 
+ * real_ptr. That is the job of the caller, if appropriate
+ */
+# define GC_MARK_FO(real_ptr, mark_proc) \
+{ \
+    (*(mark_proc))(real_ptr); \
+    while (!GC_mark_stack_empty()) MARK_FROM_MARK_STACK(); \
+    if (GC_mark_state != MS_NONE) { \
+        GC_set_mark_bit(real_ptr); \
+        while (!GC_mark_some((ptr_t)0)) {} \
+    } \
+}
+
+extern GC_bool GC_mark_stack_too_small;
+				/* We need a larger mark stack.  May be	*/
+				/* set by client supplied mark routines.*/
+
+typedef int mark_state_t;	/* Current state of marking, as follows:*/
+				/* Used to remember where we are during */
+				/* concurrent marking.			*/
+
+				/* We say something is dirty if it was	*/
+				/* written since the last time we	*/
+				/* retrieved dirty bits.  We say it's 	*/
+				/* grungy if it was marked dirty in the	*/
+				/* last set of bits we retrieved.	*/
+				
+				/* Invariant I: all roots and marked	*/
+				/* objects p are either dirty, or point */
+				/* to objects q that are either marked 	*/
+				/* or a pointer to q appears in a range	*/
+				/* on the mark stack.			*/
+
+# define MS_NONE 0		/* No marking in progress. I holds.	*/
+				/* Mark stack is empty.			*/
+
+# define MS_PUSH_RESCUERS 1	/* Rescuing objects are currently 	*/
+				/* being pushed.  I holds, except	*/
+				/* that grungy roots may point to 	*/
+				/* unmarked objects, as may marked	*/
+				/* grungy objects above scan_ptr.	*/
+
+# define MS_PUSH_UNCOLLECTABLE 2
+				/* I holds, except that marked 		*/
+				/* uncollectable objects above scan_ptr */
+				/* may point to unmarked objects.	*/
+				/* Roots may point to unmarked objects	*/
+
+# define MS_ROOTS_PUSHED 3	/* I holds, mark stack may be nonempty  */
+
+# define MS_PARTIALLY_INVALID 4	/* I may not hold, e.g. because of M.S. */
+				/* overflow.  However marked heap	*/
+				/* objects below scan_ptr point to	*/
+				/* marked or stacked objects.		*/
+
+# define MS_INVALID 5		/* I may not hold.			*/
+
+extern mark_state_t GC_mark_state;
+
+#endif  /* GC_PMARK_H */
+
diff --git a/boehm-gc/include/private/gc_priv.h b/boehm-gc/include/private/gc_priv.h
new file mode 100644
index 000000000..4dbfa7d1f
--- /dev/null
+++ b/boehm-gc/include/private/gc_priv.h
@@ -0,0 +1,2015 @@
+/* 
+ * Copyright 1988, 1989 Hans-J. Boehm, Alan J. Demers
+ * Copyright (c) 1991-1994 by Xerox Corporation.  All rights reserved.
+ * Copyright (c) 1996-1999 by Silicon Graphics.  All rights reserved.
+ * Copyright (c) 1999-2001 by Hewlett-Packard Company. All rights reserved.
+ *
+ *
+ * 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.
+ */
+ 
+
+# ifndef GC_PRIVATE_H
+# define GC_PRIVATE_H
+
+/* Autoconf definitions. */
+/* FIXME: This should really be included directly from each .c file. */
+#include <gc_config.h>
+
+#if defined(mips) && defined(SYSTYPE_BSD) && defined(sony_news)
+    /* sony RISC NEWS, NEWSOS 4 */
+#   define BSD_TIME
+/*    typedef long ptrdiff_t;   -- necessary on some really old systems	*/
+#endif
+
+#if defined(mips) && defined(SYSTYPE_BSD43)
+    /* MIPS RISCOS 4 */
+#   define BSD_TIME
+#endif
+
+#ifdef DGUX
+#   include <sys/types.h>
+#   include <sys/time.h>
+#   include <sys/resource.h>
+#endif /* DGUX */
+
+#ifdef BSD_TIME
+#   include <sys/types.h>
+#   include <sys/time.h>
+#   include <sys/resource.h>
+#endif /* BSD_TIME */
+
+# ifndef _GC_H
+#   include "../gc.h"
+# endif
+
+# ifndef GC_MARK_H
+#   include "../gc_mark.h"
+# endif
+
+typedef GC_word word;
+typedef GC_signed_word signed_word;
+
+typedef int GC_bool;
+# define TRUE 1
+# define FALSE 0
+
+typedef char * ptr_t;	/* A generic pointer to which we can add	*/
+			/* byte displacements.				*/
+			/* Preferably identical to caddr_t, if it 	*/
+			/* exists.					*/
+			
+# ifndef GCCONFIG_H
+#   include "gcconfig.h"
+# endif
+
+# ifndef HEADERS_H
+#   include "gc_hdrs.h"
+# endif
+
+#if defined(__STDC__)
+#   include <stdlib.h>
+#   if !(defined( sony_news ) )
+#       include <stddef.h>
+#   endif
+#   define VOLATILE volatile
+#else
+#   ifdef MSWIN32
+#   	include <stdlib.h>
+#   endif
+#   define VOLATILE
+#endif
+
+#if 0 /* defined(__GNUC__) doesn't work yet */
+# define EXPECT(expr, outcome) __builtin_expect(expr,outcome)
+  /* Equivalent to (expr), but predict that usually (expr)==outcome. */
+#else
+# define EXPECT(expr, outcome) (expr)
+#endif /* __GNUC__ */
+
+# ifndef GC_LOCKS_H
+#   include "gc_locks.h"
+# endif
+
+# ifdef STACK_GROWS_DOWN
+#   define COOLER_THAN >
+#   define HOTTER_THAN <
+#   define MAKE_COOLER(x,y) if ((word)(x)+(y) > (word)(x)) {(x) += (y);} \
+			    else {(x) = (word)ONES;}
+#   define MAKE_HOTTER(x,y) (x) -= (y)
+# else
+#   define COOLER_THAN <
+#   define HOTTER_THAN >
+#   define MAKE_COOLER(x,y) if ((word)(x)-(y) < (word)(x)) {(x) -= (y);} else {(x) = 0;}
+#   define MAKE_HOTTER(x,y) (x) += (y)
+# endif
+
+#if defined(AMIGA) && defined(__SASC)
+#   define GC_FAR __far
+#else
+#   define GC_FAR
+#endif
+
+
+/*********************************/
+/*                               */
+/* Definitions for conservative  */
+/* collector                     */
+/*                               */
+/*********************************/
+
+/*********************************/
+/*                               */
+/* Easily changeable parameters  */
+/*                               */
+/*********************************/
+
+/* #define STUBBORN_ALLOC */
+		    /* Enable stubborm allocation, and thus a limited	*/
+		    /* form of incremental collection w/o dirty bits.	*/
+
+/* #define ALL_INTERIOR_POINTERS */
+		    /* Forces all pointers into the interior of an 	*/
+		    /* object to be considered valid.  Also causes the	*/
+		    /* sizes of all objects to be inflated by at least 	*/
+		    /* one byte.  This should suffice to guarantee	*/
+		    /* that in the presence of a compiler that does	*/
+		    /* not perform garbage-collector-unsafe		*/
+		    /* optimizations, all portable, strictly ANSI	*/
+		    /* conforming C programs should be safely usable	*/
+		    /* with malloc replaced by GC_malloc and free	*/
+		    /* calls removed.  There are several disadvantages: */
+		    /* 1. There are probably no interesting, portable,	*/
+		    /*    strictly ANSI	conforming C programs.		*/
+		    /* 2. This option makes it hard for the collector	*/
+		    /*    to allocate space that is not ``pointed to''  */
+		    /*    by integers, etc.  Under SunOS 4.X with a 	*/
+		    /*    statically linked libc, we empiricaly		*/
+		    /*    observed that it would be difficult to 	*/
+		    /*	  allocate individual objects larger than 100K.	*/
+		    /* 	  Even if only smaller objects are allocated,	*/
+		    /*    more swap space is likely to be needed.       */
+		    /*    Fortunately, much of this will never be	*/
+		    /*    touched.					*/
+		    /* If you can easily avoid using this option, do.	*/
+		    /* If not, try to keep individual objects small.	*/
+		    /* This is now really controlled at startup,	*/
+		    /* through GC_all_interior_pointers.		*/
+		    
+#define PRINTSTATS  /* Print garbage collection statistics          	*/
+		    /* For less verbose output, undefine in reclaim.c 	*/
+
+#define PRINTTIMES  /* Print the amount of time consumed by each garbage   */
+		    /* collection.                                         */
+
+#define PRINTBLOCKS /* Print object sizes associated with heap blocks,     */
+		    /* whether the objects are atomic or composite, and    */
+		    /* whether or not the block was found to be empty      */
+		    /* during the reclaim phase.  Typically generates       */
+		    /* about one screenful per garbage collection.         */
+#undef PRINTBLOCKS
+
+#ifdef SILENT
+#  ifdef PRINTSTATS
+#    undef PRINTSTATS
+#  endif
+#  ifdef PRINTTIMES
+#    undef PRINTTIMES
+#  endif
+#  ifdef PRINTNBLOCKS
+#    undef PRINTNBLOCKS
+#  endif
+#endif
+
+#if defined(PRINTSTATS) && !defined(GATHERSTATS)
+#   define GATHERSTATS
+#endif
+
+#if defined(PRINTSTATS) || !defined(SMALL_CONFIG)
+#   define CONDPRINT  /* Print some things if GC_print_stats is set */
+#endif
+
+#define GC_INVOKE_FINALIZERS() GC_notify_or_invoke_finalizers()
+
+#define MERGE_SIZES /* Round up some object sizes, so that fewer distinct */
+		    /* free lists are actually maintained.  This applies  */
+		    /* only to the top level routines in misc.c, not to   */
+		    /* user generated code that calls GC_allocobj and     */
+		    /* GC_allocaobj directly.                             */
+		    /* Slows down average programs slightly.  May however */
+		    /* substantially reduce fragmentation if allocation   */
+		    /* request sizes are widely scattered.                */
+		    /* May save significant amounts of space for obj_map  */
+		    /* entries.						  */
+
+#if defined(USE_MARK_BYTES) && !defined(ALIGN_DOUBLE)
+#  define ALIGN_DOUBLE
+   /* We use one byte for every 2 words, which doesn't allow for	*/
+   /* odd numbered words to have mark bits.				*/
+#endif
+
+#if defined(GC_GCJ_SUPPORT) && ALIGNMENT < 8 && !defined(ALIGN_DOUBLE)
+   /* GCJ's Hashtable synchronization code requires 64-bit alignment.  */
+#  define ALIGN_DOUBLE
+#endif
+
+/* ALIGN_DOUBLE requires MERGE_SIZES at present. */
+# if defined(ALIGN_DOUBLE) && !defined(MERGE_SIZES)
+#   define MERGE_SIZES
+# endif
+
+#if !defined(DONT_ADD_BYTE_AT_END)
+# define EXTRA_BYTES GC_all_interior_pointers
+#else
+# define EXTRA_BYTES 0
+#endif
+
+
+# ifndef LARGE_CONFIG
+#   define MINHINCR 16	 /* Minimum heap increment, in blocks of HBLKSIZE  */
+			 /* Must be multiple of largest page size.	   */
+#   define MAXHINCR 2048 /* Maximum heap increment, in blocks              */
+# else
+#   define MINHINCR 64
+#   define MAXHINCR 4096
+# endif
+
+# define TIME_LIMIT 50	   /* We try to keep pause times from exceeding	 */
+			   /* this by much. In milliseconds.		 */
+
+# define BL_LIMIT GC_black_list_spacing
+			   /* If we need a block of N bytes, and we have */
+			   /* a block of N + BL_LIMIT bytes available, 	 */
+			   /* and N > BL_LIMIT,				 */
+			   /* but all possible positions in it are 	 */
+			   /* blacklisted, we just use it anyway (and	 */
+			   /* print a warning, if warnings are enabled). */
+			   /* This risks subsequently leaking the block	 */
+			   /* due to a false reference.  But not using	 */
+			   /* the block risks unreasonable immediate	 */
+			   /* heap growth.				 */
+
+/*********************************/
+/*                               */
+/* Stack saving for debugging	 */
+/*                               */
+/*********************************/
+
+#ifdef NEED_CALLINFO
+    struct callinfo {
+	word ci_pc;  	/* Caller, not callee, pc	*/
+#	if NARGS > 0
+	    word ci_arg[NARGS];	/* bit-wise complement to avoid retention */
+#	endif
+#	if defined(ALIGN_DOUBLE) && (NFRAMES * (NARGS + 1)) % 2 == 1
+	    /* Likely alignment problem. */
+	    word ci_dummy;
+#	endif
+    };
+#endif
+
+#ifdef SAVE_CALL_CHAIN
+
+/* Fill in the pc and argument information for up to NFRAMES of my	*/
+/* callers.  Ignore my frame and my callers frame.			*/
+void GC_save_callers GC_PROTO((struct callinfo info[NFRAMES]));
+  
+void GC_print_callers GC_PROTO((struct callinfo info[NFRAMES]));
+
+#endif
+
+
+/*********************************/
+/*                               */
+/* OS interface routines	 */
+/*                               */
+/*********************************/
+
+#ifdef BSD_TIME
+#   undef CLOCK_TYPE
+#   undef GET_TIME
+#   undef MS_TIME_DIFF
+#   define CLOCK_TYPE struct timeval
+#   define GET_TIME(x) { struct rusage rusage; \
+			 getrusage (RUSAGE_SELF,  &rusage); \
+			 x = rusage.ru_utime; }
+#   define MS_TIME_DIFF(a,b) ((double) (a.tv_sec - b.tv_sec) * 1000.0 \
+                               + (double) (a.tv_usec - b.tv_usec) / 1000.0)
+#else /* !BSD_TIME */
+# if defined(MSWIN32) || defined(MSWINCE)
+#   include <windows.h>
+#   include <winbase.h>
+#   define CLOCK_TYPE DWORD
+#   define GET_TIME(x) x = GetTickCount()
+#   define MS_TIME_DIFF(a,b) ((long)((a)-(b)))
+# else /* !MSWIN32, !MSWINCE, !BSD_TIME */
+#   include <time.h>
+#   if !defined(__STDC__) && defined(SPARC) && defined(SUNOS4)
+      clock_t clock();	/* Not in time.h, where it belongs	*/
+#   endif
+#   if defined(FREEBSD) && !defined(CLOCKS_PER_SEC)
+#     include <machine/limits.h>
+#     define CLOCKS_PER_SEC CLK_TCK
+#   endif
+#   if !defined(CLOCKS_PER_SEC)
+#     define CLOCKS_PER_SEC 1000000
+/*
+ * This is technically a bug in the implementation.  ANSI requires that
+ * CLOCKS_PER_SEC be defined.  But at least under SunOS4.1.1, it isn't.
+ * Also note that the combination of ANSI C and POSIX is incredibly gross
+ * here. The type clock_t is used by both clock() and times().  But on
+ * some machines these use different notions of a clock tick,  CLOCKS_PER_SEC
+ * seems to apply only to clock.  Hence we use it here.  On many machines,
+ * including SunOS, clock actually uses units of microseconds (which are
+ * not really clock ticks).
+ */
+#   endif
+#   define CLOCK_TYPE clock_t
+#   define GET_TIME(x) x = clock()
+#   define MS_TIME_DIFF(a,b) ((unsigned long) \
+		(1000.0*(double)((a)-(b))/(double)CLOCKS_PER_SEC))
+# endif /* !MSWIN32 */
+#endif /* !BSD_TIME */
+
+/* We use bzero and bcopy internally.  They may not be available.	*/
+# if defined(SPARC) && defined(SUNOS4)
+#   define BCOPY_EXISTS
+# endif
+# if defined(M68K) && defined(AMIGA)
+#   define BCOPY_EXISTS
+# endif
+# if defined(M68K) && defined(NEXT)
+#   define BCOPY_EXISTS
+# endif
+# if defined(VAX)
+#   define BCOPY_EXISTS
+# endif
+# if defined(AMIGA)
+#   include <string.h>
+#   define BCOPY_EXISTS
+# endif
+# if defined(DARWIN)
+#   include <string.h>
+#   define BCOPY_EXISTS
+# endif
+
+# ifndef BCOPY_EXISTS
+#   include <string.h>
+#   define BCOPY(x,y,n) memcpy(y, x, (size_t)(n))
+#   define BZERO(x,n)  memset(x, 0, (size_t)(n))
+# else
+#   define BCOPY(x,y,n) bcopy((char *)(x),(char *)(y),(int)(n))
+#   define BZERO(x,n) bzero((char *)(x),(int)(n))
+# endif
+
+/* Delay any interrupts or signals that may abort this thread.  Data	*/
+/* structures are in a consistent state outside this pair of calls.	*/
+/* ANSI C allows both to be empty (though the standard isn't very	*/
+/* clear on that point).  Standard malloc implementations are usually	*/
+/* neither interruptable nor thread-safe, and thus correspond to	*/
+/* empty definitions.							*/
+/* It probably doesn't make any sense to declare these to be nonempty	*/
+/* if the code is being optimized, since signal safety relies on some	*/
+/* ordering constraints that are typically not obeyed by optimizing	*/
+/* compilers.								*/
+# ifdef PCR
+#   define DISABLE_SIGNALS() \
+		 PCR_Th_SetSigMask(PCR_allSigsBlocked,&GC_old_sig_mask)
+#   define ENABLE_SIGNALS() \
+		PCR_Th_SetSigMask(&GC_old_sig_mask, NIL)
+# else
+#   if defined(THREADS) || defined(AMIGA)  \
+	|| defined(MSWIN32) || defined(MSWINCE) || defined(MACOS) \
+	|| defined(DJGPP) || defined(NO_SIGNALS) 
+			/* Also useful for debugging.		*/
+	/* Should probably use thr_sigsetmask for GC_SOLARIS_THREADS. */
+#     define DISABLE_SIGNALS()
+#     define ENABLE_SIGNALS()
+#   else
+#     define DISABLE_SIGNALS() GC_disable_signals()
+	void GC_disable_signals();
+#     define ENABLE_SIGNALS() GC_enable_signals()
+	void GC_enable_signals();
+#   endif
+# endif
+
+/*
+ * Stop and restart mutator threads.
+ */
+# ifdef PCR
+#     include "th/PCR_ThCtl.h"
+#     define STOP_WORLD() \
+ 	PCR_ThCtl_SetExclusiveMode(PCR_ThCtl_ExclusiveMode_stopNormal, \
+ 				   PCR_allSigsBlocked, \
+ 				   PCR_waitForever)
+#     define START_WORLD() \
+	PCR_ThCtl_SetExclusiveMode(PCR_ThCtl_ExclusiveMode_null, \
+ 				   PCR_allSigsBlocked, \
+ 				   PCR_waitForever);
+# else
+#   if defined(GC_SOLARIS_THREADS) || defined(GC_WIN32_THREADS) \
+	|| defined(GC_PTHREADS)
+      void GC_stop_world();
+      void GC_start_world();
+#     define STOP_WORLD() GC_stop_world()
+#     define START_WORLD() GC_start_world()
+#   else
+#     define STOP_WORLD()
+#     define START_WORLD()
+#   endif
+# endif
+
+/* Abandon ship */
+# ifdef PCR
+#   define ABORT(s) PCR_Base_Panic(s)
+# else
+#   ifdef SMALL_CONFIG
+#	define ABORT(msg) abort();
+#   else
+	GC_API void GC_abort GC_PROTO((GC_CONST char * msg));
+#       define ABORT(msg) GC_abort(msg);
+#   endif
+# endif
+
+/* Exit abnormally, but without making a mess (e.g. out of memory) */
+# ifdef PCR
+#   define EXIT() PCR_Base_Exit(1,PCR_waitForever)
+# else
+#   define EXIT() (void)exit(1)
+# endif
+
+/* Print warning message, e.g. almost out of memory.	*/
+# define WARN(msg,arg) (*GC_current_warn_proc)("GC Warning: " msg, (GC_word)(arg))
+extern GC_warn_proc GC_current_warn_proc;
+
+/* Get environment entry */
+#if !defined(NO_GETENV)
+#   if defined(EMPTY_GETENV_RESULTS)
+	/* Workaround for a reputed Wine bug.	*/
+	static inline char * fixed_getenv(const char *name)
+	{
+	  char * tmp = getenv(name);
+	  if (tmp == 0 || strlen(tmp) == 0)
+	    return 0;
+	  return tmp;
+	}
+#       define GETENV(name) fixed_getenv(name)
+#   else
+#       define GETENV(name) getenv(name)
+#   endif
+#else
+#   define GETENV(name) 0
+#endif
+
+#if defined(DARWIN)
+#      if defined(POWERPC)
+#              if CPP_WORDSZ == 32
+#                define GC_THREAD_STATE_T ppc_thread_state_t
+#		 define GC_MACH_THREAD_STATE PPC_THREAD_STATE
+#		 define GC_MACH_THREAD_STATE_COUNT PPC_THREAD_STATE_COUNT
+#                define GC_MACH_HEADER mach_header
+#                define GC_MACH_SECTION section
+#                define GC_GETSECTBYNAME getsectbynamefromheader
+#              else
+#                define GC_THREAD_STATE_T ppc_thread_state64_t
+#		 define GC_MACH_THREAD_STATE PPC_THREAD_STATE64
+#		 define GC_MACH_THREAD_STATE_COUNT PPC_THREAD_STATE64_COUNT
+#                define GC_MACH_HEADER mach_header_64
+#                define GC_MACH_SECTION section_64
+#                define GC_GETSECTBYNAME getsectbynamefromheader_64
+#              endif
+#      elif defined(I386) || defined(X86_64)
+#              if CPP_WORDSZ == 32
+#                define GC_THREAD_STATE_T x86_thread_state32_t
+#                define GC_MACH_THREAD_STATE x86_THREAD_STATE32
+#                define GC_MACH_THREAD_STATE_COUNT x86_THREAD_STATE32_COUNT
+#                define GC_MACH_HEADER mach_header
+#                define GC_MACH_SECTION section
+#                define GC_GETSECTBYNAME getsectbynamefromheader
+#              else
+#                define GC_THREAD_STATE_T x86_thread_state64_t
+#                define GC_MACH_THREAD_STATE x86_THREAD_STATE64
+#                define GC_MACH_THREAD_STATE_COUNT x86_THREAD_STATE64_COUNT
+#                define GC_MACH_HEADER mach_header_64
+#                define GC_MACH_SECTION section_64
+#                define GC_GETSECTBYNAME getsectbynamefromheader_64
+#              endif
+#      else
+#              error define GC_THREAD_STATE_T
+#              define GC_MACH_THREAD_STATE MACHINE_THREAD_STATE
+#              define GC_MACH_THREAD_STATE_COUNT MACHINE_THREAD_STATE_COUNT
+#      endif
+/* Try to work out the right way to access thread state structure members.
+   The structure has changed its definition in different Darwin versions.
+   This now defaults to the (older) names without __, thus hopefully,
+   not breaking any existing Makefile.direct builds.  */
+#      if defined (HAS_PPC_THREAD_STATE___R0) \
+	 || defined (HAS_PPC_THREAD_STATE64___R0) \
+	 || defined (HAS_X86_THREAD_STATE32___EAX) \
+	 || defined (HAS_X86_THREAD_STATE64___RAX)
+#        define THREAD_FLD(x) __ ## x
+#      else
+#        define THREAD_FLD(x) x
+#      endif
+#endif
+/*********************************/
+/*                               */
+/* Word-size-dependent defines   */
+/*                               */
+/*********************************/
+
+#if CPP_WORDSZ == 32
+#  define WORDS_TO_BYTES(x)   ((x)<<2)
+#  define BYTES_TO_WORDS(x)   ((x)>>2)
+#  define LOGWL               ((word)5)    /* log[2] of CPP_WORDSZ */
+#  define modWORDSZ(n) ((n) & 0x1f)        /* n mod size of word	    */
+#  if ALIGNMENT != 4
+#	define UNALIGNED
+#  endif
+#endif
+
+#if CPP_WORDSZ == 64
+#  define WORDS_TO_BYTES(x)   ((x)<<3)
+#  define BYTES_TO_WORDS(x)   ((x)>>3)
+#  define LOGWL               ((word)6)    /* log[2] of CPP_WORDSZ */
+#  define modWORDSZ(n) ((n) & 0x3f)        /* n mod size of word	    */
+#  if ALIGNMENT != 8
+#	define UNALIGNED
+#  endif
+#endif
+
+#define WORDSZ ((word)CPP_WORDSZ)
+#define SIGNB  ((word)1 << (WORDSZ-1))
+#define BYTES_PER_WORD      ((word)(sizeof (word)))
+#define ONES                ((word)(signed_word)(-1))
+#define divWORDSZ(n) ((n) >> LOGWL)	   /* divide n by size of word      */
+
+/*********************/
+/*                   */
+/*  Size Parameters  */
+/*                   */
+/*********************/
+
+/*  heap block size, bytes. Should be power of 2 */
+
+#ifndef HBLKSIZE
+# ifdef SMALL_CONFIG
+#   define CPP_LOG_HBLKSIZE 10
+# else
+#   if (CPP_WORDSZ == 32) || (defined(HPUX) && defined(HP_PA))
+      /* HPUX/PA seems to use 4K pages with the 64 bit ABI */
+#     define CPP_LOG_HBLKSIZE 12
+#   else
+#     define CPP_LOG_HBLKSIZE 13
+#   endif
+# endif
+#else
+# if HBLKSIZE == 512
+#   define CPP_LOG_HBLKSIZE 9
+# endif
+# if HBLKSIZE == 1024
+#   define CPP_LOG_HBLKSIZE 10
+# endif
+# if HBLKSIZE == 2048
+#   define CPP_LOG_HBLKSIZE 11
+# endif
+# if HBLKSIZE == 4096
+#   define CPP_LOG_HBLKSIZE 12
+# endif
+# if HBLKSIZE == 8192
+#   define CPP_LOG_HBLKSIZE 13
+# endif
+# if HBLKSIZE == 16384
+#   define CPP_LOG_HBLKSIZE 14
+# endif
+# ifndef CPP_LOG_HBLKSIZE
+    --> fix HBLKSIZE
+# endif
+# undef HBLKSIZE
+#endif
+# define CPP_HBLKSIZE (1 << CPP_LOG_HBLKSIZE)
+# define LOG_HBLKSIZE   ((word)CPP_LOG_HBLKSIZE)
+# define HBLKSIZE ((word)CPP_HBLKSIZE)
+
+
+/*  max size objects supported by freelist (larger objects may be   */
+/*  allocated, but less efficiently)                                */
+
+#define CPP_MAXOBJBYTES (CPP_HBLKSIZE/2)
+#define MAXOBJBYTES ((word)CPP_MAXOBJBYTES)
+#define CPP_MAXOBJSZ    BYTES_TO_WORDS(CPP_MAXOBJBYTES)
+#define MAXOBJSZ ((word)CPP_MAXOBJSZ)
+		
+# define divHBLKSZ(n) ((n) >> LOG_HBLKSIZE)
+
+# define HBLK_PTR_DIFF(p,q) divHBLKSZ((ptr_t)p - (ptr_t)q)
+	/* Equivalent to subtracting 2 hblk pointers.	*/
+	/* We do it this way because a compiler should	*/
+	/* find it hard to use an integer division	*/
+	/* instead of a shift.  The bundled SunOS 4.1	*/
+	/* o.w. sometimes pessimizes the subtraction to	*/
+	/* involve a call to .div.			*/
+ 
+# define modHBLKSZ(n) ((n) & (HBLKSIZE-1))
+ 
+# define HBLKPTR(objptr) ((struct hblk *)(((word) (objptr)) & ~(HBLKSIZE-1)))
+
+# define HBLKDISPL(objptr) (((word) (objptr)) & (HBLKSIZE-1))
+
+/* Round up byte allocation requests to integral number of words, etc. */
+# define ROUNDED_UP_WORDS(n) \
+	BYTES_TO_WORDS((n) + (WORDS_TO_BYTES(1) - 1 + EXTRA_BYTES))
+# ifdef ALIGN_DOUBLE
+#       define ALIGNED_WORDS(n) \
+	    (BYTES_TO_WORDS((n) + WORDS_TO_BYTES(2) - 1 + EXTRA_BYTES) & ~1)
+# else
+#       define ALIGNED_WORDS(n) ROUNDED_UP_WORDS(n)
+# endif
+# define SMALL_OBJ(bytes) ((bytes) <= (MAXOBJBYTES - EXTRA_BYTES))
+# define ADD_SLOP(bytes) ((bytes) + EXTRA_BYTES)
+# ifndef MIN_WORDS
+    /* MIN_WORDS is the size of the smallest allocated object.	*/
+    /* 1 and 2 are the only valid values.			*/
+    /* 2 must be used if:					*/
+    /* - GC_gcj_malloc can be used for objects of requested 	*/
+    /*   size  smaller than 2 words, or				*/
+    /* - USE_MARK_BYTES is defined.				*/
+#   if defined(USE_MARK_BYTES) || defined(GC_GCJ_SUPPORT)
+#     define MIN_WORDS 2   	/* Smallest allocated object.	*/
+#   else
+#     define MIN_WORDS 1
+#   endif
+# endif
+
+
+/*
+ * Hash table representation of sets of pages.  This assumes it is
+ * OK to add spurious entries to sets.
+ * Used by black-listing code, and perhaps by dirty bit maintenance code.
+ */
+ 
+# ifdef LARGE_CONFIG
+#   define LOG_PHT_ENTRIES  20  /* Collisions likely at 1M blocks,	*/
+				/* which is >= 4GB.  Each table takes	*/
+				/* 128KB, some of which may never be	*/
+				/* touched.				*/
+# else
+#   ifdef SMALL_CONFIG
+#     define LOG_PHT_ENTRIES  14 /* Collisions are likely if heap grows	*/
+				 /* to more than 16K hblks = 64MB.	*/
+				 /* Each hash table occupies 2K bytes.   */
+#   else /* default "medium" configuration */
+#     define LOG_PHT_ENTRIES  16 /* Collisions are likely if heap grows	*/
+				 /* to more than 64K hblks >= 256MB.	*/
+				 /* Each hash table occupies 8K bytes.  */
+				 /* Even for somewhat smaller heaps, 	*/
+				 /* say half that, collisions may be an	*/
+				 /* issue because we blacklist 		*/
+				 /* addresses outside the heap.		*/
+#   endif
+# endif
+# define PHT_ENTRIES ((word)1 << LOG_PHT_ENTRIES)
+# define PHT_SIZE (PHT_ENTRIES >> LOGWL)
+typedef word page_hash_table[PHT_SIZE];
+
+# define PHT_HASH(addr) ((((word)(addr)) >> LOG_HBLKSIZE) & (PHT_ENTRIES - 1))
+
+# define get_pht_entry_from_index(bl, index) \
+		(((bl)[divWORDSZ(index)] >> modWORDSZ(index)) & 1)
+# define set_pht_entry_from_index(bl, index) \
+		(bl)[divWORDSZ(index)] |= (word)1 << modWORDSZ(index)
+# define clear_pht_entry_from_index(bl, index) \
+		(bl)[divWORDSZ(index)] &= ~((word)1 << modWORDSZ(index))
+/* And a dumb but thread-safe version of set_pht_entry_from_index.	*/
+/* This sets (many) extra bits.						*/
+# define set_pht_entry_from_index_safe(bl, index) \
+		(bl)[divWORDSZ(index)] = ONES
+	
+
+
+/********************************************/
+/*                                          */
+/*    H e a p   B l o c k s                 */
+/*                                          */
+/********************************************/
+
+/*  heap block header */
+#define HBLKMASK   (HBLKSIZE-1)
+
+#define BITS_PER_HBLK (CPP_HBLKSIZE * 8)
+
+#define MARK_BITS_PER_HBLK (BITS_PER_HBLK/CPP_WORDSZ)
+	   /* upper bound                                    */
+	   /* We allocate 1 bit/word, unless USE_MARK_BYTES  */
+	   /* is defined.  Only the first word   	     */
+	   /* in each object is actually marked.             */
+
+# ifdef USE_MARK_BYTES
+#   define MARK_BITS_SZ (MARK_BITS_PER_HBLK/2)
+	/* Unlike the other case, this is in units of bytes.		*/
+	/* We actually allocate only every second mark bit, since we	*/
+	/* force all objects to be doubleword aligned.			*/
+	/* However, each mark bit is allocated as a byte.		*/
+# else
+#   define MARK_BITS_SZ (MARK_BITS_PER_HBLK/CPP_WORDSZ)
+# endif
+
+/* We maintain layout maps for heap blocks containing objects of a given */
+/* size.  Each entry in this map describes a byte offset and has the	 */
+/* following type.							 */
+typedef unsigned char map_entry_type;
+
+struct hblkhdr {
+    word hb_sz;  /* If in use, size in words, of objects in the block. */
+		 /* if free, the size in bytes of the whole block      */
+    struct hblk * hb_next; 	/* Link field for hblk free list	 */
+    				/* and for lists of chunks waiting to be */
+    				/* reclaimed.				 */
+    struct hblk * hb_prev;	/* Backwards link for free list.	*/
+    word hb_descr;   		/* object descriptor for marking.  See	*/
+    				/* mark.h.				*/
+    map_entry_type * hb_map;	
+    			/* A pointer to a pointer validity map of the block. */
+    		      	/* See GC_obj_map.				     */
+    		     	/* Valid for all blocks with headers.		     */
+    		     	/* Free blocks point to GC_invalid_map.		     */
+    unsigned char hb_obj_kind;
+    			 /* Kind of objects in the block.  Each kind 	*/
+    			 /* identifies a mark procedure and a set of 	*/
+    			 /* list headers.  Sometimes called regions.	*/
+    unsigned char hb_flags;
+#	define IGNORE_OFF_PAGE	1	/* Ignore pointers that do not	*/
+					/* point to the first page of 	*/
+					/* this object.			*/
+#	define WAS_UNMAPPED 2	/* This is a free block, which has	*/
+				/* been unmapped from the address 	*/
+				/* space.				*/
+				/* GC_remap must be invoked on it	*/
+				/* before it can be reallocated.	*/
+				/* Only set with USE_MUNMAP.		*/
+    unsigned short hb_last_reclaimed;
+    				/* Value of GC_gc_no when block was	*/
+    				/* last allocated or swept. May wrap.   */
+				/* For a free block, this is maintained */
+				/* only for USE_MUNMAP, and indicates	*/
+				/* when the header was allocated, or	*/
+				/* when the size of the block last	*/
+				/* changed.				*/
+#   ifdef USE_MARK_BYTES
+      union {
+        char _hb_marks[MARK_BITS_SZ];
+			    /* The i'th byte is 1 if the object 	*/
+			    /* starting at word 2i is marked, 0 o.w.	*/
+	word dummy;	/* Force word alignment of mark bytes. */
+      } _mark_byte_union;
+#     define hb_marks _mark_byte_union._hb_marks
+#   else
+      word hb_marks[MARK_BITS_SZ];
+			    /* Bit i in the array refers to the             */
+			    /* object starting at the ith word (header      */
+			    /* INCLUDED) in the heap block.                 */
+			    /* The lsb of word 0 is numbered 0.		    */
+			    /* Unused bits are invalid, and are 	    */
+			    /* occasionally set, e.g for uncollectable	    */
+			    /* objects.					    */
+#   endif /* !USE_MARK_BYTES */
+};
+
+/*  heap block body */
+
+# define BODY_SZ (HBLKSIZE/sizeof(word))
+
+struct hblk {
+    word hb_body[BODY_SZ];
+};
+
+# define HBLK_IS_FREE(hdr) ((hdr) -> hb_map == GC_invalid_map)
+
+# define OBJ_SZ_TO_BLOCKS(sz) \
+    divHBLKSZ(WORDS_TO_BYTES(sz) + HBLKSIZE-1)
+    /* Size of block (in units of HBLKSIZE) needed to hold objects of	*/
+    /* given sz (in words).						*/
+
+/* Object free list link */
+# define obj_link(p) (*(ptr_t *)(p))
+
+# define LOG_MAX_MARK_PROCS 6
+# define MAX_MARK_PROCS (1 << LOG_MAX_MARK_PROCS)
+
+/* Root sets.  Logically private to mark_rts.c.  But we don't want the	*/
+/* tables scanned, so we put them here.					*/
+/* MAX_ROOT_SETS is the maximum number of ranges that can be 	*/
+/* registered as static roots. 					*/
+# ifdef LARGE_CONFIG
+#   define MAX_ROOT_SETS 4096
+# else
+    /* GCJ LOCAL: MAX_ROOT_SETS increased to permit more shared */
+    /* libraries to be loaded.                                  */ 
+#   define MAX_ROOT_SETS 1024
+# endif
+
+# define MAX_EXCLUSIONS (MAX_ROOT_SETS/4)
+/* Maximum number of segments that can be excluded from root sets.	*/
+
+/*
+ * Data structure for excluded static roots.
+ */
+struct exclusion {
+    ptr_t e_start;
+    ptr_t e_end;
+};
+
+/* Data structure for list of root sets.				*/
+/* We keep a hash table, so that we can filter out duplicate additions.	*/
+/* Under Win32, we need to do a better job of filtering overlaps, so	*/
+/* we resort to sequential search, and pay the price.			*/
+struct roots {
+	ptr_t r_start;
+	ptr_t r_end;
+#	if !defined(MSWIN32) && !defined(MSWINCE) && !defined(CYGWIN32)
+	  struct roots * r_next;
+#	endif
+	GC_bool r_tmp;
+	  	/* Delete before registering new dynamic libraries */
+};
+
+#if !defined(MSWIN32) && !defined(MSWINCE) && !defined(CYGWIN32)
+    /* Size of hash table index to roots.	*/
+#   define LOG_RT_SIZE 6
+#   define RT_SIZE (1 << LOG_RT_SIZE) /* Power of 2, may be != MAX_ROOT_SETS */
+#endif
+
+/* Lists of all heap blocks and free lists	*/
+/* as well as other random data structures	*/
+/* that should not be scanned by the		*/
+/* collector.					*/
+/* These are grouped together in a struct	*/
+/* so that they can be easily skipped by the	*/
+/* GC_mark routine.				*/
+/* The ordering is weird to make GC_malloc	*/
+/* faster by keeping the important fields	*/
+/* sufficiently close together that a		*/
+/* single load of a base register will do.	*/
+/* Scalars that could easily appear to		*/
+/* be pointers are also put here.		*/
+/* The main fields should precede any 		*/
+/* conditionally included fields, so that	*/
+/* gc_inl.h will work even if a different set	*/
+/* of macros is defined when the client is	*/
+/* compiled.					*/
+
+struct _GC_arrays {
+  word _heapsize;
+  word _max_heapsize;
+  word _requested_heapsize;	/* Heap size due to explicit expansion */
+  ptr_t _last_heap_addr;
+  ptr_t _prev_heap_addr;
+  word _large_free_bytes;
+	/* Total bytes contained in blocks on large object free */
+	/* list.						*/
+  word _large_allocd_bytes;
+  	/* Total number of bytes in allocated large objects blocks.	*/
+  	/* For the purposes of this counter and the next one only, a 	*/
+  	/* large object is one that occupies a block of at least	*/
+  	/* 2*HBLKSIZE.							*/
+  word _max_large_allocd_bytes;
+  	/* Maximum number of bytes that were ever allocated in		*/
+  	/* large object blocks.  This is used to help decide when it	*/
+  	/* is safe to split up a large block.				*/
+  word _words_allocd_before_gc;
+		/* Number of words allocated before this	*/
+		/* collection cycle.				*/
+# ifndef SEPARATE_GLOBALS
+    word _words_allocd;
+  	/* Number of words allocated during this collection cycle */
+# endif
+  word _words_wasted;
+  	/* Number of words wasted due to internal fragmentation	*/
+  	/* in large objects, or due to dropping blacklisted     */
+	/* blocks, since last gc.  Approximate.                 */
+  word _words_finalized;
+  	/* Approximate number of words in objects (and headers)	*/
+  	/* That became ready for finalization in the last 	*/
+  	/* collection.						*/
+  word _non_gc_bytes_at_gc;
+  	/* Number of explicitly managed bytes of storage 	*/
+  	/* at last collection.					*/
+  word _mem_freed;
+  	/* Number of explicitly deallocated words of memory	*/
+  	/* since last collection.				*/
+  word _finalizer_mem_freed;
+  	/* Words of memory explicitly deallocated while 	*/
+  	/* finalizers were running.  Used to approximate mem.	*/
+  	/* explicitly deallocated by finalizers.		*/
+  ptr_t _scratch_end_ptr;
+  ptr_t _scratch_last_end_ptr;
+	/* Used by headers.c, and can easily appear to point to	*/
+	/* heap.						*/
+  GC_mark_proc _mark_procs[MAX_MARK_PROCS];
+  	/* Table of user-defined mark procedures.  There is	*/
+	/* a small number of these, which can be referenced	*/
+	/* by DS_PROC mark descriptors.  See gc_mark.h.		*/
+
+# ifndef SEPARATE_GLOBALS
+    ptr_t _objfreelist[MAXOBJSZ+1];
+			  /* free list for objects */
+    ptr_t _aobjfreelist[MAXOBJSZ+1];
+			  /* free list for atomic objs 	*/
+# endif
+
+  ptr_t _uobjfreelist[MAXOBJSZ+1];
+			  /* uncollectable but traced objs 	*/
+			  /* objects on this and auobjfreelist  */
+			  /* are always marked, except during   */
+			  /* garbage collections.		*/
+# ifdef ATOMIC_UNCOLLECTABLE
+    ptr_t _auobjfreelist[MAXOBJSZ+1];
+# endif
+			  /* uncollectable but traced objs 	*/
+
+# ifdef GATHERSTATS
+    word _composite_in_use;
+   		/* Number of words in accessible composite	*/
+		/* objects.					*/
+    word _atomic_in_use;
+   		/* Number of words in accessible atomic		*/
+		/* objects.					*/
+# endif
+# ifdef USE_MUNMAP
+    word _unmapped_bytes;
+# endif
+# ifdef MERGE_SIZES
+    unsigned _size_map[WORDS_TO_BYTES(MAXOBJSZ+1)];
+    	/* Number of words to allocate for a given allocation request in */
+    	/* bytes.							 */
+# endif 
+
+# ifdef STUBBORN_ALLOC
+    ptr_t _sobjfreelist[MAXOBJSZ+1];
+# endif
+  			  /* free list for immutable objects	*/
+  map_entry_type * _obj_map[MAXOBJSZ+1];
+                       /* If not NIL, then a pointer to a map of valid  */
+    		       /* object addresses. _obj_map[sz][i] is j if the	*/
+    		       /* address block_start+i is a valid pointer      */
+    		       /* to an object at block_start +			*/
+ 		       /* WORDS_TO_BYTES(BYTES_TO_WORDS(i) - j)		*/
+  		       /* I.e. j is a word displacement from the	*/
+  		       /* object beginning.				*/
+  		       /* The entry is OBJ_INVALID if the corresponding	*/
+  		       /* address is not a valid pointer.  It is 	*/
+  		       /* OFFSET_TOO_BIG if the value j would be too 	*/
+  		       /* large to fit in the entry.  (Note that the	*/
+  		       /* size of these entries matters, both for 	*/
+  		       /* space consumption and for cache utilization.)	*/
+#   define OFFSET_TOO_BIG 0xfe
+#   define OBJ_INVALID 0xff
+#   define MAP_ENTRY(map, bytes) (map)[bytes]
+#   define MAP_ENTRIES HBLKSIZE
+#   define MAP_SIZE MAP_ENTRIES
+#   define CPP_MAX_OFFSET (OFFSET_TOO_BIG - 1)	
+#   define MAX_OFFSET ((word)CPP_MAX_OFFSET)
+    /* The following are used only if GC_all_interior_ptrs != 0 */
+# 	define VALID_OFFSET_SZ \
+	  (CPP_MAX_OFFSET > WORDS_TO_BYTES(CPP_MAXOBJSZ)? \
+	   CPP_MAX_OFFSET+1 \
+	   : WORDS_TO_BYTES(CPP_MAXOBJSZ)+1)
+  	char _valid_offsets[VALID_OFFSET_SZ];
+				/* GC_valid_offsets[i] == TRUE ==> i 	*/
+				/* is registered as a displacement.	*/
+  	char _modws_valid_offsets[sizeof(word)];
+				/* GC_valid_offsets[i] ==>		  */
+				/* GC_modws_valid_offsets[i%sizeof(word)] */
+#   define OFFSET_VALID(displ) \
+	  (GC_all_interior_pointers || GC_valid_offsets[displ])
+# ifdef STUBBORN_ALLOC
+    page_hash_table _changed_pages;
+        /* Stubborn object pages that were changes since last call to	*/
+	/* GC_read_changed.						*/
+    page_hash_table _prev_changed_pages;
+        /* Stubborn object pages that were changes before last call to	*/
+	/* GC_read_changed.						*/
+# endif
+# if defined(PROC_VDB) || defined(MPROTECT_VDB)
+    page_hash_table _grungy_pages; /* Pages that were dirty at last 	   */
+				     /* GC_read_dirty.			   */
+# endif
+# ifdef MPROTECT_VDB
+    VOLATILE page_hash_table _dirty_pages;	
+			/* Pages dirtied since last GC_read_dirty. */
+# endif
+# ifdef PROC_VDB
+    page_hash_table _written_pages;	/* Pages ever dirtied	*/
+# endif
+# ifdef LARGE_CONFIG
+#   if CPP_WORDSZ > 32
+#     define MAX_HEAP_SECTS 4096 	/* overflows at roughly 64 GB	   */
+#   else
+#     define MAX_HEAP_SECTS 768		/* Separately added heap sections. */
+#   endif
+# else
+#   ifdef SMALL_CONFIG
+#     define MAX_HEAP_SECTS 128		/* Roughly 256MB (128*2048*1K)	*/
+#   else
+#     define MAX_HEAP_SECTS 384		/* Roughly 3GB			*/
+#   endif
+# endif
+  struct HeapSect {
+      ptr_t hs_start; word hs_bytes;
+  } _heap_sects[MAX_HEAP_SECTS];
+# if defined(MSWIN32) || defined(MSWINCE) || defined(CYGWIN32)
+    ptr_t _heap_bases[MAX_HEAP_SECTS];
+    		/* Start address of memory regions obtained from kernel. */
+# endif
+# ifdef MSWINCE
+    word _heap_lengths[MAX_HEAP_SECTS];
+    		/* Commited lengths of memory regions obtained from kernel. */
+# endif
+  struct roots _static_roots[MAX_ROOT_SETS];
+# if !defined(MSWIN32) && !defined(MSWINCE) && !defined(CYGWIN32)
+    struct roots * _root_index[RT_SIZE];
+# endif
+  struct exclusion _excl_table[MAX_EXCLUSIONS];
+  /* Block header index; see gc_headers.h */
+  bottom_index * _all_nils;
+  bottom_index * _top_index [TOP_SZ];
+#ifdef SAVE_CALL_CHAIN
+  struct callinfo _last_stack[NFRAMES];	/* Stack at last garbage collection.*/
+  					/* Useful for debugging	mysterious  */
+  					/* object disappearances.	    */
+  					/* In the multithreaded case, we    */
+  					/* currently only save the calling  */
+  					/* stack.			    */
+#endif
+};
+
+GC_API GC_FAR struct _GC_arrays GC_arrays; 
+
+# ifndef SEPARATE_GLOBALS
+#   define GC_objfreelist GC_arrays._objfreelist
+#   define GC_aobjfreelist GC_arrays._aobjfreelist
+#   define GC_words_allocd GC_arrays._words_allocd
+# endif
+# define GC_uobjfreelist GC_arrays._uobjfreelist
+# ifdef ATOMIC_UNCOLLECTABLE
+#   define GC_auobjfreelist GC_arrays._auobjfreelist
+# endif
+# define GC_sobjfreelist GC_arrays._sobjfreelist
+# define GC_valid_offsets GC_arrays._valid_offsets
+# define GC_modws_valid_offsets GC_arrays._modws_valid_offsets
+# ifdef STUBBORN_ALLOC
+#    define GC_changed_pages GC_arrays._changed_pages
+#    define GC_prev_changed_pages GC_arrays._prev_changed_pages
+# endif
+# define GC_obj_map GC_arrays._obj_map
+# define GC_last_heap_addr GC_arrays._last_heap_addr
+# define GC_prev_heap_addr GC_arrays._prev_heap_addr
+# define GC_words_wasted GC_arrays._words_wasted
+# define GC_large_free_bytes GC_arrays._large_free_bytes
+# define GC_large_allocd_bytes GC_arrays._large_allocd_bytes
+# define GC_max_large_allocd_bytes GC_arrays._max_large_allocd_bytes
+# define GC_words_finalized GC_arrays._words_finalized
+# define GC_non_gc_bytes_at_gc GC_arrays._non_gc_bytes_at_gc
+# define GC_mem_freed GC_arrays._mem_freed
+# define GC_finalizer_mem_freed GC_arrays._finalizer_mem_freed
+# define GC_scratch_end_ptr GC_arrays._scratch_end_ptr
+# define GC_scratch_last_end_ptr GC_arrays._scratch_last_end_ptr
+# define GC_mark_procs GC_arrays._mark_procs
+# define GC_heapsize GC_arrays._heapsize
+# define GC_max_heapsize GC_arrays._max_heapsize
+# define GC_requested_heapsize GC_arrays._requested_heapsize
+# define GC_words_allocd_before_gc GC_arrays._words_allocd_before_gc
+# define GC_heap_sects GC_arrays._heap_sects
+# define GC_last_stack GC_arrays._last_stack
+# ifdef USE_MUNMAP
+#   define GC_unmapped_bytes GC_arrays._unmapped_bytes
+# endif
+# if defined(MSWIN32) || defined(MSWINCE) || defined (CYGWIN32)
+#   define GC_heap_bases GC_arrays._heap_bases
+# endif
+# ifdef MSWINCE
+#   define GC_heap_lengths GC_arrays._heap_lengths
+# endif
+# define GC_static_roots GC_arrays._static_roots
+# define GC_root_index GC_arrays._root_index
+# define GC_excl_table GC_arrays._excl_table
+# define GC_all_nils GC_arrays._all_nils
+# define GC_top_index GC_arrays._top_index
+# if defined(PROC_VDB) || defined(MPROTECT_VDB)
+#   define GC_grungy_pages GC_arrays._grungy_pages
+# endif
+# ifdef MPROTECT_VDB
+#   define GC_dirty_pages GC_arrays._dirty_pages
+# endif
+# ifdef PROC_VDB
+#   define GC_written_pages GC_arrays._written_pages
+# endif
+# ifdef GATHERSTATS
+#   define GC_composite_in_use GC_arrays._composite_in_use
+#   define GC_atomic_in_use GC_arrays._atomic_in_use
+# endif
+# ifdef MERGE_SIZES
+#   define GC_size_map GC_arrays._size_map
+# endif
+
+# define beginGC_arrays ((ptr_t)(&GC_arrays))
+# define endGC_arrays (((ptr_t)(&GC_arrays)) + (sizeof GC_arrays))
+
+#define USED_HEAP_SIZE (GC_heapsize - GC_large_free_bytes)
+
+/* Object kinds: */
+# define MAXOBJKINDS 16
+
+extern struct obj_kind {
+   ptr_t *ok_freelist;	/* Array of free listheaders for this kind of object */
+   			/* Point either to GC_arrays or to storage allocated */
+   			/* with GC_scratch_alloc.			     */
+   struct hblk **ok_reclaim_list;
+   			/* List headers for lists of blocks waiting to be */
+   			/* swept.					  */
+   word ok_descriptor;  /* Descriptor template for objects in this	*/
+   			/* block.					*/
+   GC_bool ok_relocate_descr;
+   			/* Add object size in bytes to descriptor 	*/
+   			/* template to obtain descriptor.  Otherwise	*/
+   			/* template is used as is.			*/
+   GC_bool ok_init;   /* Clear objects before putting them on the free list. */
+} GC_obj_kinds[MAXOBJKINDS];
+
+# define beginGC_obj_kinds ((ptr_t)(&GC_obj_kinds))
+# define endGC_obj_kinds (beginGC_obj_kinds + (sizeof GC_obj_kinds))
+
+/* Variables that used to be in GC_arrays, but need to be accessed by 	*/
+/* inline allocation code.  If they were in GC_arrays, the inlined 	*/
+/* allocation code would include GC_arrays offsets (as it did), which	*/
+/* introduce maintenance problems.					*/
+
+#ifdef SEPARATE_GLOBALS
+  word GC_words_allocd;
+  	/* Number of words allocated during this collection cycle */
+  ptr_t GC_objfreelist[MAXOBJSZ+1];
+			  /* free list for NORMAL objects */
+# define beginGC_objfreelist ((ptr_t)(&GC_objfreelist))
+# define endGC_objfreelist (beginGC_objfreelist + sizeof(GC_objfreelist))
+
+  ptr_t GC_aobjfreelist[MAXOBJSZ+1];
+			  /* free list for atomic (PTRFREE) objs 	*/
+# define beginGC_aobjfreelist ((ptr_t)(&GC_aobjfreelist))
+# define endGC_aobjfreelist (beginGC_aobjfreelist + sizeof(GC_aobjfreelist))
+#endif
+
+/* Predefined kinds: */
+# define PTRFREE 0
+# define NORMAL  1
+# define UNCOLLECTABLE 2
+# ifdef ATOMIC_UNCOLLECTABLE
+#   define AUNCOLLECTABLE 3
+#   define STUBBORN 4
+#   define IS_UNCOLLECTABLE(k) (((k) & ~1) == UNCOLLECTABLE)
+# else
+#   define STUBBORN 3
+#   define IS_UNCOLLECTABLE(k) ((k) == UNCOLLECTABLE)
+# endif
+
+extern int GC_n_kinds;
+
+GC_API word GC_fo_entries;
+
+extern word GC_n_heap_sects;	/* Number of separately added heap	*/
+				/* sections.				*/
+
+extern word GC_page_size;
+
+# if defined(MSWIN32) || defined(MSWINCE) || defined(CYGWIN32)
+  struct _SYSTEM_INFO;
+  extern struct _SYSTEM_INFO GC_sysinfo;
+  extern word GC_n_heap_bases;	/* See GC_heap_bases.	*/
+# endif
+
+extern word GC_total_stack_black_listed;
+			/* Number of bytes on stack blacklist. 	*/
+
+extern word GC_black_list_spacing;
+			/* Average number of bytes between blacklisted	*/
+			/* blocks. Approximate.				*/
+			/* Counts only blocks that are 			*/
+			/* "stack-blacklisted", i.e. that are 		*/
+			/* problematic in the interior of an object.	*/
+
+extern map_entry_type * GC_invalid_map;
+			/* Pointer to the nowhere valid hblk map */
+			/* Blocks pointing to this map are free. */
+
+extern struct hblk * GC_hblkfreelist[];
+				/* List of completely empty heap blocks	*/
+				/* Linked through hb_next field of 	*/
+				/* header structure associated with	*/
+				/* block.				*/
+
+extern GC_bool GC_objects_are_marked;	/* There are marked objects in  */
+					/* the heap.			*/
+
+#ifndef SMALL_CONFIG
+  extern GC_bool GC_incremental;
+			/* Using incremental/generational collection. */
+# define TRUE_INCREMENTAL \
+	(GC_incremental && GC_time_limit != GC_TIME_UNLIMITED)
+	/* True incremental, not just generational, mode */
+#else
+# define GC_incremental FALSE
+			/* Hopefully allow optimizer to remove some code. */
+# define TRUE_INCREMENTAL FALSE
+#endif
+
+extern GC_bool GC_dirty_maintained;
+				/* Dirty bits are being maintained, 	*/
+				/* either for incremental collection,	*/
+				/* or to limit the root set.		*/
+
+extern word GC_root_size;	/* Total size of registered root sections */
+
+extern GC_bool GC_debugging_started;	/* GC_debug_malloc has been called. */ 
+
+extern long GC_large_alloc_warn_interval;
+	/* Interval between unsuppressed warnings.	*/
+
+extern long GC_large_alloc_warn_suppressed;
+	/* Number of warnings suppressed so far.	*/
+
+#ifdef THREADS
+  extern GC_bool GC_world_stopped;
+#endif
+
+/* Operations */
+# ifndef abs
+#   define abs(x)  ((x) < 0? (-(x)) : (x))
+# endif
+
+
+/*  Marks are in a reserved area in                          */
+/*  each heap block.  Each word has one mark bit associated  */
+/*  with it. Only those corresponding to the beginning of an */
+/*  object are used.                                         */
+
+/* Set mark bit correctly, even if mark bits may be concurrently 	*/
+/* accessed.								*/
+#ifdef PARALLEL_MARK
+# define OR_WORD(addr, bits) \
+	{ word old; \
+	  do { \
+	    old = *((volatile word *)addr); \
+	  } while (!GC_compare_and_exchange((addr), old, old | (bits))); \
+	}
+# define OR_WORD_EXIT_IF_SET(addr, bits, exit_label) \
+	{ word old; \
+	  word my_bits = (bits); \
+	  do { \
+	    old = *((volatile word *)addr); \
+	    if (old & my_bits) goto exit_label; \
+	  } while (!GC_compare_and_exchange((addr), old, old | my_bits)); \
+	}
+#else
+# define OR_WORD(addr, bits) *(addr) |= (bits)
+# define OR_WORD_EXIT_IF_SET(addr, bits, exit_label) \
+	{ \
+	  word old = *(addr); \
+	  word my_bits = (bits); \
+	  if (old & my_bits) goto exit_label; \
+	  *(addr) = (old | my_bits); \
+	}
+#endif
+
+/* Mark bit operations */
+
+/*
+ * Retrieve, set, clear the mark bit corresponding
+ * to the nth word in a given heap block.
+ *
+ * (Recall that bit n corresponds to object beginning at word n
+ * relative to the beginning of the block, including unused words)
+ */
+
+#ifdef USE_MARK_BYTES
+# define mark_bit_from_hdr(hhdr,n) ((hhdr)->hb_marks[(n) >> 1])
+# define set_mark_bit_from_hdr(hhdr,n) ((hhdr)->hb_marks[(n)>>1]) = 1
+# define clear_mark_bit_from_hdr(hhdr,n) ((hhdr)->hb_marks[(n)>>1]) = 0
+#else /* !USE_MARK_BYTES */
+# define mark_bit_from_hdr(hhdr,n) (((hhdr)->hb_marks[divWORDSZ(n)] \
+			    >> (modWORDSZ(n))) & (word)1)
+# define set_mark_bit_from_hdr(hhdr,n) \
+			    OR_WORD((hhdr)->hb_marks+divWORDSZ(n), \
+				    (word)1 << modWORDSZ(n))
+# define clear_mark_bit_from_hdr(hhdr,n) (hhdr)->hb_marks[divWORDSZ(n)] \
+				&= ~((word)1 << modWORDSZ(n))
+#endif /* !USE_MARK_BYTES */
+
+/* Important internal collector routines */
+
+ptr_t GC_approx_sp GC_PROTO((void));
+  
+GC_bool GC_should_collect GC_PROTO((void));
+  
+void GC_apply_to_all_blocks GC_PROTO(( \
+    void (*fn) GC_PROTO((struct hblk *h, word client_data)), \
+    word client_data));
+  			/* Invoke fn(hbp, client_data) for each 	*/
+  			/* allocated heap block.			*/
+struct hblk * GC_next_used_block GC_PROTO((struct hblk * h));
+  			/* Return first in-use block >= h	*/
+struct hblk * GC_prev_block GC_PROTO((struct hblk * h));
+  			/* Return last block <= h.  Returned block	*/
+  			/* is managed by GC, but may or may not be in	*/
+			/* use.						*/
+void GC_mark_init GC_PROTO((void));
+void GC_clear_marks GC_PROTO((void));	/* Clear mark bits for all heap objects. */
+void GC_invalidate_mark_state GC_PROTO((void));
+					/* Tell the marker that	marked 	   */
+  					/* objects may point to	unmarked   */
+  					/* ones, and roots may point to	   */
+  					/* unmarked objects.		   */
+  					/* Reset mark stack.		   */
+GC_bool GC_mark_stack_empty GC_PROTO((void));
+GC_bool GC_mark_some GC_PROTO((ptr_t cold_gc_frame));
+  			/* Perform about one pages worth of marking	*/
+  			/* work of whatever kind is needed.  Returns	*/
+  			/* quickly if no collection is in progress.	*/
+  			/* Return TRUE if mark phase finished.		*/
+void GC_initiate_gc GC_PROTO((void));
+				/* initiate collection.			*/
+  				/* If the mark state is invalid, this	*/
+  				/* becomes full colleection.  Otherwise */
+  				/* it's partial.			*/
+void GC_push_all GC_PROTO((ptr_t bottom, ptr_t top));
+				/* Push everything in a range 		*/
+  				/* onto mark stack.			*/
+void GC_push_selected GC_PROTO(( \
+    ptr_t bottom, \
+    ptr_t top, \
+    int (*dirty_fn) GC_PROTO((struct hblk *h)), \
+    void (*push_fn) GC_PROTO((ptr_t bottom, ptr_t top)) ));
+				  /* Push all pages h in [b,t) s.t. 	*/
+				  /* select_fn(h) != 0 onto mark stack. */
+#ifndef SMALL_CONFIG
+  void GC_push_conditional GC_PROTO((ptr_t b, ptr_t t, GC_bool all));
+#else
+# define GC_push_conditional(b, t, all) GC_push_all(b, t)
+#endif
+                                /* Do either of the above, depending	*/
+				/* on the third arg.			*/
+void GC_push_all_stack GC_PROTO((ptr_t b, ptr_t t));
+				    /* As above, but consider		*/
+				    /*  interior pointers as valid  	*/
+void GC_push_all_eager GC_PROTO((ptr_t b, ptr_t t));
+				    /* Same as GC_push_all_stack, but   */
+				    /* ensures that stack is scanned	*/
+				    /* immediately, not just scheduled  */
+				    /* for scanning.			*/
+#ifndef THREADS
+  void GC_push_all_stack_partially_eager GC_PROTO(( \
+      ptr_t bottom, ptr_t top, ptr_t cold_gc_frame ));
+			/* Similar to GC_push_all_eager, but only the	*/
+			/* part hotter than cold_gc_frame is scanned	*/
+			/* immediately.  Needed to ensure that callee-	*/
+			/* save registers are not missed.		*/
+#else
+  /* In the threads case, we push part of the current thread stack	*/
+  /* with GC_push_all_eager when we push the registers.  This gets the  */
+  /* callee-save registers that may disappear.  The remainder of the	*/
+  /* stacks are scheduled for scanning in *GC_push_other_roots, which	*/
+  /* is thread-package-specific.					*/
+#endif
+void GC_push_current_stack GC_PROTO((ptr_t cold_gc_frame));
+  			/* Push enough of the current stack eagerly to	*/
+  			/* ensure that callee-save registers saved in	*/
+  			/* GC frames are scanned.			*/
+  			/* In the non-threads case, schedule entire	*/
+  			/* stack for scanning.				*/
+void GC_push_roots GC_PROTO((GC_bool all, ptr_t cold_gc_frame));
+  			/* Push all or dirty roots.	*/
+extern void (*GC_push_other_roots) GC_PROTO((void));
+  			/* Push system or application specific roots	*/
+  			/* onto the mark stack.  In some environments	*/
+  			/* (e.g. threads environments) this is		*/
+  			/* predfined to be non-zero.  A client supplied */
+  			/* replacement should also call the original	*/
+  			/* function.					*/
+extern void GC_push_gc_structures GC_PROTO((void));
+			/* Push GC internal roots.  These are normally	*/
+			/* included in the static data segment, and 	*/
+			/* Thus implicitly pushed.  But we must do this	*/
+			/* explicitly if normal root processing is 	*/
+			/* disabled.  Calls the following:		*/
+	extern void GC_push_finalizer_structures GC_PROTO((void));
+	extern void GC_push_stubborn_structures GC_PROTO((void));
+#	ifdef THREADS
+	  extern void GC_push_thread_structures GC_PROTO((void));
+#	endif
+extern void (*GC_start_call_back) GC_PROTO((void));
+  			/* Called at start of full collections.		*/
+  			/* Not called if 0.  Called with allocation 	*/
+  			/* lock held.					*/
+  			/* 0 by default.				*/
+# if defined(USE_GENERIC_PUSH_REGS)
+  void GC_generic_push_regs GC_PROTO((ptr_t cold_gc_frame));
+# else
+  void GC_push_regs GC_PROTO((void));
+# endif
+# if defined(SPARC) || defined(IA64)
+  /* Cause all stacked registers to be saved in memory.  Return a	*/
+  /* pointer to the top of the corresponding memory stack.		*/
+  word GC_save_regs_in_stack GC_PROTO((void));
+# endif
+			/* Push register contents onto mark stack.	*/
+  			/* If NURSERY is defined, the default push	*/
+  			/* action can be overridden with GC_push_proc	*/
+
+# ifdef NURSERY
+    extern void (*GC_push_proc)(ptr_t);
+# endif
+# if defined(MSWIN32) || defined(MSWINCE)
+  void __cdecl GC_push_one GC_PROTO((word p));
+# else
+  void GC_push_one GC_PROTO((word p));
+			      /* If p points to an object, mark it    */
+                              /* and push contents on the mark stack  */
+  			      /* Pointer recognition test always      */
+  			      /* accepts interior pointers, i.e. this */
+  			      /* is appropriate for pointers found on */
+  			      /* stack.				      */
+# endif
+# if defined(PRINT_BLACK_LIST) || defined(KEEP_BACK_PTRS)
+  void GC_mark_and_push_stack GC_PROTO((word p, ptr_t source));
+				/* Ditto, omits plausibility test	*/
+# else
+  void GC_mark_and_push_stack GC_PROTO((word p));
+# endif
+void GC_push_marked GC_PROTO((struct hblk * h, hdr * hhdr));
+		/* Push contents of all marked objects in h onto	*/
+		/* mark stack.						*/
+#ifdef SMALL_CONFIG
+# define GC_push_next_marked_dirty(h) GC_push_next_marked(h)
+#else
+  struct hblk * GC_push_next_marked_dirty GC_PROTO((struct hblk * h));
+		/* Invoke GC_push_marked on next dirty block above h.	*/
+		/* Return a pointer just past the end of this block.	*/
+#endif /* !SMALL_CONFIG */
+struct hblk * GC_push_next_marked GC_PROTO((struct hblk * h));
+  		/* Ditto, but also mark from clean pages.	*/
+struct hblk * GC_push_next_marked_uncollectable GC_PROTO((struct hblk * h));
+  		/* Ditto, but mark only from uncollectable pages.	*/
+GC_bool GC_stopped_mark GC_PROTO((GC_stop_func stop_func));
+ 			/* Stop world and mark from all roots	*/
+  			/* and rescuers.			*/
+void GC_clear_hdr_marks GC_PROTO((hdr * hhdr));
+				    /* Clear the mark bits in a header */
+void GC_set_hdr_marks GC_PROTO((hdr * hhdr));
+ 				    /* Set the mark bits in a header */
+void GC_set_fl_marks GC_PROTO((ptr_t p));
+				    /* Set all mark bits associated with */
+				    /* a free list.			 */
+void GC_add_roots_inner GC_PROTO((char * b, char * e, GC_bool tmp));
+void GC_remove_roots_inner GC_PROTO((char * b, char * e));
+GC_bool GC_is_static_root GC_PROTO((ptr_t p));
+  		/* Is the address p in one of the registered static	*/
+  		/* root sections?					*/
+# if defined(MSWIN32) || defined(_WIN32_WCE_EMULATION) || defined(CYGWIN32)
+GC_bool GC_is_tmp_root GC_PROTO((ptr_t p));
+		/* Is the address p in one of the temporary static	*/
+		/* root sections?					*/
+# endif
+void GC_register_dynamic_libraries GC_PROTO((void));
+  		/* Add dynamic library data sections to the root set. */
+
+GC_bool GC_register_main_static_data GC_PROTO((void));
+		/* We need to register the main data segment.  Returns	*/
+		/* TRUE unless this is done implicitly as part of	*/
+		/* dynamic library registration.			*/
+  
+/* Machine dependent startup routines */
+ptr_t GC_get_stack_base GC_PROTO((void));	/* Cold end of stack */
+#ifdef IA64
+  ptr_t GC_get_register_stack_base GC_PROTO((void));
+  					/* Cold end of register stack.	*/
+#endif
+void GC_register_data_segments GC_PROTO((void));
+  
+/* Black listing: */
+void GC_bl_init GC_PROTO((void));
+# ifdef PRINT_BLACK_LIST
+      void GC_add_to_black_list_normal GC_PROTO((word p, ptr_t source));
+			/* Register bits as a possible future false	*/
+			/* reference from the heap or static data	*/
+#     define GC_ADD_TO_BLACK_LIST_NORMAL(bits, source) \
+      		if (GC_all_interior_pointers) { \
+		  GC_add_to_black_list_stack(bits, (ptr_t)(source)); \
+		} else { \
+  		  GC_add_to_black_list_normal(bits, (ptr_t)(source)); \
+		}
+# else
+      void GC_add_to_black_list_normal GC_PROTO((word p));
+#     define GC_ADD_TO_BLACK_LIST_NORMAL(bits, source) \
+      		if (GC_all_interior_pointers) { \
+		  GC_add_to_black_list_stack(bits); \
+		} else { \
+  		  GC_add_to_black_list_normal(bits); \
+		}
+# endif
+
+# ifdef PRINT_BLACK_LIST
+    void GC_add_to_black_list_stack GC_PROTO((word p, ptr_t source));
+# else
+    void GC_add_to_black_list_stack GC_PROTO((word p));
+# endif
+struct hblk * GC_is_black_listed GC_PROTO((struct hblk * h, word len));
+  			/* If there are likely to be false references	*/
+  			/* to a block starting at h of the indicated    */
+  			/* length, then return the next plausible	*/
+  			/* starting location for h that might avoid	*/
+  			/* these false references.			*/
+void GC_promote_black_lists GC_PROTO((void));
+  			/* Declare an end to a black listing phase.	*/
+void GC_unpromote_black_lists GC_PROTO((void));
+  			/* Approximately undo the effect of the above.	*/
+  			/* This actually loses some information, but	*/
+  			/* only in a reasonably safe way.		*/
+word GC_number_stack_black_listed GC_PROTO(( \
+	struct hblk *start, struct hblk *endp1));
+  			/* Return the number of (stack) blacklisted	*/
+  			/* blocks in the range for statistical		*/
+  			/* purposes.					*/
+  		 	
+ptr_t GC_scratch_alloc GC_PROTO((word bytes));
+  				/* GC internal memory allocation for	*/
+  				/* small objects.  Deallocation is not  */
+  				/* possible.				*/
+  	
+/* Heap block layout maps: */			
+void GC_invalidate_map GC_PROTO((hdr * hhdr));
+  				/* Remove the object map associated	*/
+  				/* with the block.  This identifies	*/
+  				/* the block as invalid to the mark	*/
+  				/* routines.				*/
+GC_bool GC_add_map_entry GC_PROTO((word sz));
+  				/* Add a heap block map for objects of	*/
+  				/* size sz to obj_map.			*/
+  				/* Return FALSE on failure.		*/
+void GC_register_displacement_inner GC_PROTO((word offset));
+  				/* Version of GC_register_displacement	*/
+  				/* that assumes lock is already held	*/
+  				/* and signals are already disabled.	*/
+  
+/*  hblk allocation: */		
+void GC_new_hblk GC_PROTO((word size_in_words, int kind));
+  				/* Allocate a new heap block, and build */
+  				/* a free list in it.			*/				
+
+ptr_t GC_build_fl GC_PROTO((struct hblk *h, word sz,
+			   GC_bool clear,  ptr_t list));
+				/* Build a free list for objects of 	*/
+				/* size sz in block h.  Append list to	*/
+				/* end of the free lists.  Possibly	*/
+				/* clear objects on the list.  Normally	*/
+				/* called by GC_new_hblk, but also	*/
+				/* called explicitly without GC lock.	*/
+
+struct hblk * GC_allochblk GC_PROTO(( \
+	word size_in_words, int kind, unsigned flags));
+				/* Allocate a heap block, inform	*/
+				/* the marker that block is valid	*/
+				/* for objects of indicated size.	*/
+
+ptr_t GC_alloc_large GC_PROTO((word lw, int k, unsigned flags));
+			/* Allocate a large block of size lw words.	*/
+			/* The block is not cleared.			*/
+			/* Flags is 0 or IGNORE_OFF_PAGE.		*/
+			/* Calls GC_allchblk to do the actual 		*/
+			/* allocation, but also triggers GC and/or	*/
+			/* heap expansion as appropriate.		*/
+			/* Does not update GC_words_allocd, but does	*/
+			/* other accounting.				*/
+
+ptr_t GC_alloc_large_and_clear GC_PROTO((word lw, int k, unsigned flags));
+			/* As above, but clear block if appropriate	*/
+			/* for kind k.					*/
+
+void GC_freehblk GC_PROTO((struct hblk * p));
+				/* Deallocate a heap block and mark it  */
+  				/* as invalid.				*/
+  				
+/*  Misc GC: */
+void GC_init_inner GC_PROTO((void));
+GC_bool GC_expand_hp_inner GC_PROTO((word n));
+void GC_start_reclaim GC_PROTO((int abort_if_found));
+  				/* Restore unmarked objects to free	*/
+  				/* lists, or (if abort_if_found is	*/
+  				/* TRUE) report them.			*/
+  				/* Sweeping of small object pages is	*/
+  				/* largely deferred.			*/
+void GC_continue_reclaim GC_PROTO((word sz, int kind));
+  				/* Sweep pages of the given size and	*/
+  				/* kind, as long as possible, and	*/
+  				/* as long as the corr. free list is    */
+  				/* empty.				*/
+void GC_reclaim_or_delete_all GC_PROTO((void));
+  				/* Arrange for all reclaim lists to be	*/
+  				/* empty.  Judiciously choose between	*/
+  				/* sweeping and discarding each page.	*/
+GC_bool GC_reclaim_all GC_PROTO((GC_stop_func stop_func, GC_bool ignore_old));
+  				/* Reclaim all blocks.  Abort (in a	*/
+  				/* consistent state) if f returns TRUE. */
+GC_bool GC_block_empty GC_PROTO((hdr * hhdr));
+ 				/* Block completely unmarked? 	*/
+GC_bool GC_never_stop_func GC_PROTO((void));
+				/* Returns FALSE.		*/
+GC_bool GC_try_to_collect_inner GC_PROTO((GC_stop_func f));
+
+				/* Collect; caller must have acquired	*/
+				/* lock and disabled signals.		*/
+				/* Collection is aborted if f returns	*/
+				/* TRUE.  Returns TRUE if it completes	*/
+				/* successfully.			*/
+# define GC_gcollect_inner() \
+	(void) GC_try_to_collect_inner(GC_never_stop_func)
+void GC_finish_collection GC_PROTO((void));
+ 				/* Finish collection.  Mark bits are	*/
+  				/* consistent and lock is still held.	*/
+GC_bool GC_collect_or_expand GC_PROTO(( \
+	word needed_blocks, GC_bool ignore_off_page));
+  				/* Collect or expand heap in an attempt */
+  				/* make the indicated number of free	*/
+  				/* blocks available.  Should be called	*/
+  				/* until the blocks are available or	*/
+  				/* until it fails by returning FALSE.	*/
+
+extern GC_bool GC_is_initialized;	/* GC_init() has been run.	*/
+
+#if defined(MSWIN32) || defined(MSWINCE)
+  void GC_deinit GC_PROTO((void));
+                                /* Free any resources allocated by      */
+                                /* GC_init                              */
+#endif
+
+void GC_collect_a_little_inner GC_PROTO((int n));
+  				/* Do n units worth of garbage 		*/
+  				/* collection work, if appropriate.	*/
+  				/* A unit is an amount appropriate for  */
+  				/* HBLKSIZE bytes of allocation.	*/
+/* ptr_t GC_generic_malloc GC_PROTO((word lb, int k)); */
+  				/* Allocate an object of the given	*/
+  				/* kind.  By default, there are only	*/
+  				/* a few kinds: composite(pointerfree), */
+				/* atomic, uncollectable, etc.		*/
+				/* We claim it's possible for clever	*/
+				/* client code that understands GC	*/
+				/* internals to add more, e.g. to	*/
+				/* communicate object layout info	*/
+				/* to the collector.			*/
+				/* The actual decl is in gc_mark.h.	*/
+ptr_t GC_generic_malloc_ignore_off_page GC_PROTO((size_t b, int k));
+  				/* As above, but pointers past the 	*/
+  				/* first page of the resulting object	*/
+  				/* are ignored.				*/
+ptr_t GC_generic_malloc_inner GC_PROTO((word lb, int k));
+  				/* Ditto, but I already hold lock, etc.	*/
+ptr_t GC_generic_malloc_words_small_inner GC_PROTO((word lw, int k));
+				/* Analogous to the above, but assumes	*/
+				/* a small object size, and bypasses	*/
+				/* MERGE_SIZES mechanism.		*/
+ptr_t GC_generic_malloc_words_small GC_PROTO((size_t lw, int k));
+  				/* As above, but size in units of words */
+  				/* Bypasses MERGE_SIZES.  Assumes	*/
+  				/* words <= MAXOBJSZ.			*/
+ptr_t GC_generic_malloc_inner_ignore_off_page GC_PROTO((size_t lb, int k));
+  				/* Allocate an object, where		*/
+  				/* the client guarantees that there	*/
+  				/* will always be a pointer to the 	*/
+  				/* beginning of the object while the	*/
+  				/* object is live.			*/
+ptr_t GC_allocobj GC_PROTO((word sz, int kind));
+  				/* Make the indicated 			*/
+  				/* free list nonempty, and return its	*/
+  				/* head.				*/
+
+void GC_free_inner(GC_PTR p);
+  
+void GC_init_headers GC_PROTO((void));
+struct hblkhdr * GC_install_header GC_PROTO((struct hblk *h));
+  				/* Install a header for block h.	*/
+  				/* Return 0 on failure, or the header	*/
+  				/* otherwise.				*/
+GC_bool GC_install_counts GC_PROTO((struct hblk * h, word sz));
+  				/* Set up forwarding counts for block	*/
+  				/* h of size sz.			*/
+  				/* Return FALSE on failure.		*/
+void GC_remove_header GC_PROTO((struct hblk * h));
+  				/* Remove the header for block h.	*/
+void GC_remove_counts GC_PROTO((struct hblk * h, word sz));
+  				/* Remove forwarding counts for h.	*/
+hdr * GC_find_header GC_PROTO((ptr_t h)); /* Debugging only.		*/
+  
+void GC_finalize GC_PROTO((void));
+ 			/* Perform all indicated finalization actions	*/
+  			/* on unmarked objects.				*/
+  			/* Unreachable finalizable objects are enqueued	*/
+  			/* for processing by GC_invoke_finalizers.	*/
+  			/* Invoked with lock.				*/
+
+void GC_notify_or_invoke_finalizers GC_PROTO((void));
+			/* If GC_finalize_on_demand is not set, invoke	*/
+			/* eligible finalizers. Otherwise:		*/
+			/* Call *GC_finalizer_notifier if there are	*/
+			/* finalizers to be run, and we haven't called	*/
+			/* this procedure yet this GC cycle.		*/
+
+GC_API GC_PTR GC_make_closure GC_PROTO((GC_finalization_proc fn, GC_PTR data));
+GC_API void GC_debug_invoke_finalizer GC_PROTO((GC_PTR obj, GC_PTR data));
+			/* Auxiliary fns to make finalization work	*/
+			/* correctly with displaced pointers introduced	*/
+			/* by the debugging allocators.			*/
+  			
+void GC_add_to_heap GC_PROTO((struct hblk *p, word bytes));
+  			/* Add a HBLKSIZE aligned chunk to the heap.	*/
+  
+void GC_print_obj GC_PROTO((ptr_t p));
+  			/* P points to somewhere inside an object with	*/
+  			/* debugging info.  Print a human readable	*/
+  			/* description of the object to stderr.		*/
+extern void (*GC_check_heap) GC_PROTO((void));
+  			/* Check that all objects in the heap with 	*/
+  			/* debugging info are intact.  			*/
+  			/* Add any that are not to GC_smashed list.	*/
+extern void (*GC_print_all_smashed) GC_PROTO((void));
+			/* Print GC_smashed if it's not empty.		*/
+			/* Clear GC_smashed list.			*/
+extern void GC_print_all_errors GC_PROTO((void));
+			/* Print smashed and leaked objects, if any.	*/
+			/* Clear the lists of such objects.		*/
+extern void (*GC_print_heap_obj) GC_PROTO((ptr_t p));
+  			/* If possible print s followed by a more	*/
+  			/* detailed description of the object 		*/
+  			/* referred to by p.				*/
+#if defined(LINUX) && defined(__ELF__) && !defined(SMALL_CONFIG)
+  void GC_print_address_map GC_PROTO((void));
+  			/* Print an address map of the process.		*/
+#endif
+
+extern GC_bool GC_have_errors;  /* We saw a smashed or leaked object.	*/
+				/* Call error printing routine 		*/
+				/* occasionally.			*/
+extern GC_bool GC_print_stats;	/* Produce at least some logging output	*/
+				/* Set from environment variable.	*/
+
+#ifndef NO_DEBUGGING
+  extern GC_bool GC_dump_regularly;  /* Generate regular debugging dumps. */
+# define COND_DUMP if (GC_dump_regularly) GC_dump();
+#else
+# define COND_DUMP
+#endif
+
+#ifdef KEEP_BACK_PTRS
+  extern long GC_backtraces;
+  void GC_generate_random_backtrace_no_gc(void);
+#endif
+
+extern GC_bool GC_print_back_height;
+
+#ifdef MAKE_BACK_GRAPH
+  void GC_print_back_graph_stats(void);
+#endif
+
+/* Macros used for collector internal allocation.	*/
+/* These assume the collector lock is held.		*/
+#ifdef DBG_HDRS_ALL
+    extern GC_PTR GC_debug_generic_malloc_inner(size_t lb, int k);
+    extern GC_PTR GC_debug_generic_malloc_inner_ignore_off_page(size_t lb,
+								int k);
+#   define GC_INTERNAL_MALLOC GC_debug_generic_malloc_inner
+#   define GC_INTERNAL_MALLOC_IGNORE_OFF_PAGE \
+		 GC_debug_generic_malloc_inner_ignore_off_page
+#   ifdef THREADS
+#       define GC_INTERNAL_FREE GC_debug_free_inner
+#   else
+#       define GC_INTERNAL_FREE GC_debug_free
+#   endif
+#else
+#   define GC_INTERNAL_MALLOC GC_generic_malloc_inner
+#   define GC_INTERNAL_MALLOC_IGNORE_OFF_PAGE \
+		 GC_generic_malloc_inner_ignore_off_page
+#   ifdef THREADS
+#       define GC_INTERNAL_FREE GC_free_inner
+#   else
+#       define GC_INTERNAL_FREE GC_free
+#   endif
+#endif
+
+/* Memory unmapping: */
+#ifdef USE_MUNMAP
+  void GC_unmap_old(void);
+  void GC_merge_unmapped(void);
+  void GC_unmap(ptr_t start, word bytes);
+  void GC_remap(ptr_t start, word bytes);
+  void GC_unmap_gap(ptr_t start1, word bytes1, ptr_t start2, word bytes2);
+#endif
+
+/* Virtual dirty bit implementation:		*/
+/* Each implementation exports the following:	*/
+void GC_read_dirty GC_PROTO((void));
+			/* Retrieve dirty bits.	*/
+GC_bool GC_page_was_dirty GC_PROTO((struct hblk *h));
+  			/* Read retrieved dirty bits.	*/
+GC_bool GC_page_was_ever_dirty GC_PROTO((struct hblk *h));
+  			/* Could the page contain valid heap pointers?	*/
+void GC_is_fresh GC_PROTO((struct hblk *h, word n));
+  			/* Assert the region currently contains no	*/
+  			/* valid pointers.				*/
+void GC_remove_protection GC_PROTO((struct hblk *h, word nblocks,
+			   	    GC_bool pointerfree));
+  			/* h is about to be writteni or allocated.  Ensure  */
+			/* that it's not write protected by the virtual	    */
+			/* dirty bit implementation.			    */
+			
+void GC_dirty_init GC_PROTO((void));
+  
+/* Slow/general mark bit manipulation: */
+GC_API GC_bool GC_is_marked GC_PROTO((ptr_t p));
+void GC_clear_mark_bit GC_PROTO((ptr_t p));
+void GC_set_mark_bit GC_PROTO((ptr_t p));
+  
+/* Stubborn objects: */
+void GC_read_changed GC_PROTO((void));	/* Analogous to GC_read_dirty */
+GC_bool GC_page_was_changed GC_PROTO((struct hblk * h));
+ 				/* Analogous to GC_page_was_dirty */
+void GC_clean_changing_list GC_PROTO((void));
+ 				/* Collect obsolete changing list entries */
+void GC_stubborn_init GC_PROTO((void));
+  
+/* Debugging print routines: */
+void GC_print_block_list GC_PROTO((void));
+void GC_print_hblkfreelist GC_PROTO((void));
+void GC_print_heap_sects GC_PROTO((void));
+void GC_print_static_roots GC_PROTO((void));
+void GC_print_finalization_stats GC_PROTO((void));
+void GC_dump GC_PROTO((void));
+
+#ifdef KEEP_BACK_PTRS
+   void GC_store_back_pointer(ptr_t source, ptr_t dest);
+   void GC_marked_for_finalization(ptr_t dest);
+#  define GC_STORE_BACK_PTR(source, dest) GC_store_back_pointer(source, dest)
+#  define GC_MARKED_FOR_FINALIZATION(dest) GC_marked_for_finalization(dest)
+#else
+#  define GC_STORE_BACK_PTR(source, dest) 
+#  define GC_MARKED_FOR_FINALIZATION(dest)
+#endif
+
+/* Make arguments appear live to compiler */
+# ifdef __WATCOMC__
+    void GC_noop(void*, ...);
+# else
+#   ifdef __DMC__
+      GC_API void GC_noop(...);
+#   else
+      GC_API void GC_noop();
+#   endif
+# endif
+
+void GC_noop1 GC_PROTO((word));
+
+/* Logging and diagnostic output: 	*/
+GC_API void GC_printf GC_PROTO((GC_CONST char * format, long, long, long, long, long, long));
+			/* A version of printf that doesn't allocate,	*/
+			/* is restricted to long arguments, and		*/
+			/* (unfortunately) doesn't use varargs for	*/
+			/* portability.  Restricted to 6 args and	*/
+			/* 1K total output length.			*/
+			/* (We use sprintf.  Hopefully that doesn't	*/
+			/* allocate for long arguments.)  		*/
+# define GC_printf0(f) GC_printf(f, 0l, 0l, 0l, 0l, 0l, 0l)
+# define GC_printf1(f,a) GC_printf(f, (long)a, 0l, 0l, 0l, 0l, 0l)
+# define GC_printf2(f,a,b) GC_printf(f, (long)a, (long)b, 0l, 0l, 0l, 0l)
+# define GC_printf3(f,a,b,c) GC_printf(f, (long)a, (long)b, (long)c, 0l, 0l, 0l)
+# define GC_printf4(f,a,b,c,d) GC_printf(f, (long)a, (long)b, (long)c, \
+					    (long)d, 0l, 0l)
+# define GC_printf5(f,a,b,c,d,e) GC_printf(f, (long)a, (long)b, (long)c, \
+					      (long)d, (long)e, 0l)
+# define GC_printf6(f,a,b,c,d,e,g) GC_printf(f, (long)a, (long)b, (long)c, \
+						(long)d, (long)e, (long)g)
+
+GC_API void GC_err_printf GC_PROTO((GC_CONST char * format, long, long, long, long, long, long));
+# define GC_err_printf0(f) GC_err_puts(f)
+# define GC_err_printf1(f,a) GC_err_printf(f, (long)a, 0l, 0l, 0l, 0l, 0l)
+# define GC_err_printf2(f,a,b) GC_err_printf(f, (long)a, (long)b, 0l, 0l, 0l, 0l)
+# define GC_err_printf3(f,a,b,c) GC_err_printf(f, (long)a, (long)b, (long)c, \
+						  0l, 0l, 0l)
+# define GC_err_printf4(f,a,b,c,d) GC_err_printf(f, (long)a, (long)b, \
+						    (long)c, (long)d, 0l, 0l)
+# define GC_err_printf5(f,a,b,c,d,e) GC_err_printf(f, (long)a, (long)b, \
+						      (long)c, (long)d, \
+						      (long)e, 0l)
+# define GC_err_printf6(f,a,b,c,d,e,g) GC_err_printf(f, (long)a, (long)b, \
+							(long)c, (long)d, \
+							(long)e, (long)g)
+			/* Ditto, writes to stderr.			*/
+			
+void GC_err_puts GC_PROTO((GC_CONST char *s));
+			/* Write s to stderr, don't buffer, don't add	*/
+			/* newlines, don't ...				*/
+
+#if defined(LINUX) && !defined(SMALL_CONFIG)
+  void GC_err_write GC_PROTO((GC_CONST char *buf, size_t len));
+ 			/* Write buf to stderr, don't buffer, don't add	*/
+  			/* newlines, don't ...				*/
+#endif
+
+
+# ifdef GC_ASSERTIONS
+#	define GC_ASSERT(expr) if(!(expr)) {\
+		GC_err_printf2("Assertion failure: %s:%ld\n", \
+				__FILE__, (unsigned long)__LINE__); \
+		ABORT("assertion failure"); }
+# else 
+#	define GC_ASSERT(expr)
+# endif
+
+/* Check a compile time assertion at compile time.  The error	*/
+/* message for failure is a bit baroque, but ...		*/
+#if defined(mips) && !defined(__GNUC__)
+/* DOB: MIPSPro C gets an internal error taking the sizeof an array type. 
+   This code works correctly (ugliness is to avoid "unused var" warnings) */
+# define GC_STATIC_ASSERT(expr) do { if (0) { char j[(expr)? 1 : -1]; j[0]='\0'; j[0]=j[0]; } } while(0)
+#else
+# define GC_STATIC_ASSERT(expr) sizeof(char[(expr)? 1 : -1])
+#endif
+
+# if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
+    /* We need additional synchronization facilities from the thread	*/
+    /* support.  We believe these are less performance critical		*/
+    /* than the main garbage collector lock; standard pthreads-based	*/
+    /* implementations should be sufficient.				*/
+
+    /* The mark lock and condition variable.  If the GC lock is also 	*/
+    /* acquired, the GC lock must be acquired first.  The mark lock is	*/
+    /* used to both protect some variables used by the parallel		*/
+    /* marker, and to protect GC_fl_builder_count, below.		*/
+    /* GC_notify_all_marker() is called when				*/ 
+    /* the state of the parallel marker changes				*/
+    /* in some significant way (see gc_mark.h for details).  The	*/
+    /* latter set of events includes incrementing GC_mark_no.		*/
+    /* GC_notify_all_builder() is called when GC_fl_builder_count	*/
+    /* reaches 0.							*/
+
+     extern void GC_acquire_mark_lock();
+     extern void GC_release_mark_lock();
+     extern void GC_notify_all_builder();
+     /* extern void GC_wait_builder(); */
+     extern void GC_wait_for_reclaim();
+
+     extern word GC_fl_builder_count;	/* Protected by mark lock.	*/
+# endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
+# ifdef PARALLEL_MARK
+     extern void GC_notify_all_marker();
+     extern void GC_wait_marker();
+     extern word GC_mark_no;		/* Protected by mark lock.	*/
+
+     extern void GC_help_marker(word my_mark_no);
+		/* Try to help out parallel marker for mark cycle 	*/
+		/* my_mark_no.  Returns if the mark cycle finishes or	*/
+		/* was already done, or there was nothing to do for	*/
+		/* some other reason.					*/
+# endif /* PARALLEL_MARK */
+
+# if defined(GC_PTHREADS) && !defined(GC_SOLARIS_THREADS)
+  /* We define the thread suspension signal here, so that we can refer	*/
+  /* to it in the dirty bit implementation, if necessary.  Ideally we	*/
+  /* would allocate a (real-time ?) signal using the standard mechanism.*/
+  /* unfortunately, there is no standard mechanism.  (There is one 	*/
+  /* in Linux glibc, but it's not exported.)  Thus we continue to use	*/
+  /* the same hard-coded signals we've always used.			*/
+#  if !defined(SIG_SUSPEND)
+#   if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
+#    if defined(SPARC) && !defined(SIGPWR)
+       /* SPARC/Linux doesn't properly define SIGPWR in <signal.h>.
+        * It is aliased to SIGLOST in asm/signal.h, though.		*/
+#      define SIG_SUSPEND SIGLOST
+#    else
+       /* Linuxthreads itself uses SIGUSR1 and SIGUSR2.			*/
+#      define SIG_SUSPEND SIGPWR
+#    endif
+#   else  /* !GC_LINUX_THREADS */
+#     if defined(_SIGRTMIN)
+#       define SIG_SUSPEND _SIGRTMIN + 6
+#     else
+#       define SIG_SUSPEND SIGRTMIN + 6
+#     endif       
+#   endif
+#  endif /* !SIG_SUSPEND */
+  
+# endif
+
+# endif /* GC_PRIVATE_H */
diff --git a/boehm-gc/include/private/gcconfig.h b/boehm-gc/include/private/gcconfig.h
new file mode 100644
index 000000000..f2b568135
--- /dev/null
+++ b/boehm-gc/include/private/gcconfig.h
@@ -0,0 +1,2430 @@
+/* 
+ * Copyright 1988, 1989 Hans-J. Boehm, Alan J. Demers
+ * Copyright (c) 1991-1994 by Xerox Corporation.  All rights reserved.
+ * Copyright (c) 1996 by Silicon Graphics.  All rights reserved.
+ * Copyright (c) 2000-2004 Hewlett-Packard Development Company, L.P.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
+ * OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program
+ * for any purpose,  provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is granted,
+ * provided the above notices are retained, and a notice that the code was
+ * modified is included with the above copyright notice.
+ */
+
+/*
+ * This header is private to the gc.  It is almost always included from
+ * gc_priv.h.  However it is possible to include it by itself if just the
+ * configuration macros are needed.  In that
+ * case, a few declarations relying on types declared in gc_priv.h will be
+ * omitted.
+ */
+ 
+#ifndef GCCONFIG_H
+
+# define GCCONFIG_H
+
+# ifndef GC_PRIVATE_H
+    /* Fake ptr_t declaration, just to avoid compilation errors.	*/
+    /* This avoids many instances if "ifndef GC_PRIVATE_H" below.	*/
+    typedef struct GC_undefined_struct * ptr_t;
+# endif
+
+/* Machine dependent parameters.  Some tuning parameters can be found	*/
+/* near the top of gc_private.h.					*/
+
+/* Machine specific parts contributed by various people.  See README file. */
+
+/* First a unified test for Linux: */
+# if defined(linux) || defined(__linux__)
+#  ifndef LINUX
+#    define LINUX
+#  endif
+# endif
+
+/* And one for NetBSD: */
+# if defined(__NetBSD__)
+#    define NETBSD
+# endif
+
+/* And one for OpenBSD: */
+# if defined(__OpenBSD__)
+#    define OPENBSD
+# endif
+
+/* And one for FreeBSD: */
+# if ( defined(__FreeBSD__) || defined(__FreeBSD_kernel__) ) && !defined(FREEBSD)
+#    define FREEBSD
+# endif
+
+/* Determine the machine type: */
+# if defined(__arm__) || defined(__thumb__)
+#    define ARM32
+#    if !defined(LINUX) && !defined(NETBSD)
+#      define NOSYS
+#      define mach_type_known
+#    endif
+# endif
+# if defined(sun) && defined(mc68000)
+#    define M68K
+#    define SUNOS4
+#    define mach_type_known
+# endif
+# if defined(hp9000s300)
+#    define M68K
+#    define HP
+#    define mach_type_known
+# endif
+# if defined(OPENBSD) && defined(m68k)
+#    define M68K
+#    define mach_type_known
+# endif
+# if defined(OPENBSD) && defined(__sparc__)
+#    define SPARC
+#    define mach_type_known
+# endif
+# if defined(NETBSD) && (defined(m68k) || defined(__m68k__))
+#    define M68K
+#    define mach_type_known
+# endif
+# if defined(NETBSD) && defined(__powerpc__)
+#    define POWERPC
+#    define mach_type_known
+# endif
+# if defined(NETBSD) && (defined(__arm32__) || defined(__arm__))
+#    define ARM32
+#    define mach_type_known
+# endif
+# if defined(NETBSD) && defined(__sh__)
+#    define SH
+#    define mach_type_known
+# endif
+# if defined(vax)
+#    define VAX
+#    ifdef ultrix
+#	define ULTRIX
+#    else
+#	define BSD
+#    endif
+#    define mach_type_known
+# endif
+# if defined(__NetBSD__) && defined(__vax__)
+#    define VAX
+#    define mach_type_known
+# endif
+# if defined(mips) || defined(__mips) || defined(_mips)
+#    define MIPS
+#    if defined(nec_ews) || defined(_nec_ews)
+#      define EWS4800
+#    endif
+#    if !defined(LINUX) && !defined(EWS4800) && !defined(NETBSD)
+#      if defined(ultrix) || defined(__ultrix)
+#	 define ULTRIX
+#      else
+#	 if defined(_SYSTYPE_SVR4) || defined(SYSTYPE_SVR4) \
+	    || defined(__SYSTYPE_SVR4__)
+#	   define IRIX5   /* or IRIX 6.X */
+#	 else
+#	   define RISCOS  /* or IRIX 4.X */
+#	 endif
+#      endif
+#    endif /* !LINUX */
+#    if defined(__NetBSD__) && defined(__MIPSEL__)
+#      undef ULTRIX
+#    endif
+#    define mach_type_known
+# endif
+# if defined(DGUX) && (defined(i386) || defined(__i386__))
+#    define I386
+#    ifndef _USING_DGUX
+#    define _USING_DGUX
+#    endif
+#    define mach_type_known
+# endif
+# if defined(sequent) && (defined(i386) || defined(__i386__))
+#    define I386
+#    define SEQUENT
+#    define mach_type_known
+# endif
+# if defined(sun) && (defined(i386) || defined(__i386__))
+#    define I386
+#    define SUNOS5
+#    define mach_type_known
+# endif
+# if defined(sun) && defined(__amd64)
+#    define X86_64
+#    define SUNOS5
+#    define mach_type_known
+# endif
+# if (defined(__OS2__) || defined(__EMX__)) && defined(__32BIT__)
+#    define I386
+#    define OS2
+#    define mach_type_known
+# endif
+# if defined(ibm032)
+#   define RT
+#   define mach_type_known
+# endif
+# if defined(sun) && (defined(sparc) || defined(__sparc))
+#   define SPARC
+    /* Test for SunOS 5.x */
+#     include <errno.h>
+#     ifdef ECHRNG
+#       define SUNOS5
+#     else
+#	define SUNOS4
+#     endif
+#   define mach_type_known
+# endif
+# if defined(sparc) && defined(unix) && !defined(sun) && !defined(linux) \
+     && !defined(__OpenBSD__) && !defined(__NetBSD__) && !defined(__FreeBSD__)
+#   define SPARC
+#   define DRSNX
+#   define mach_type_known
+# endif
+# if defined(_IBMR2)
+#   define RS6000
+#   define mach_type_known
+# endif
+# if defined(__NetBSD__) && defined(__sparc__)
+#   define SPARC
+#   define mach_type_known
+# endif
+# if defined(_M_XENIX) && defined(_M_SYSV) && defined(_M_I386)
+	/* The above test may need refinement	*/
+#   define I386
+#   if defined(_SCO_ELF)
+#     define SCO_ELF
+#   else
+#     define SCO
+#   endif
+#   define mach_type_known
+# endif
+# if defined(_AUX_SOURCE)
+#   define M68K
+#   define SYSV
+#   define mach_type_known
+# endif
+# if defined(_PA_RISC1_0) || defined(_PA_RISC1_1) || defined(_PA_RISC2_0) \
+     || defined(hppa) || defined(__hppa__)
+#   define HP_PA
+#   if !defined(LINUX) && !defined(HPUX)
+#     define HPUX
+#   endif
+#   define mach_type_known
+# endif
+# if defined(__ia64) && defined(_HPUX_SOURCE)
+#   define IA64
+#   ifndef HPUX
+#     define HPUX
+#   endif
+#   define mach_type_known
+# endif
+# if defined(__BEOS__) && defined(_X86_)
+#    define I386
+#    define BEOS
+#    define mach_type_known
+# endif
+# if defined(LINUX) && (defined(i386) || defined(__i386__))
+#    define I386
+#    define mach_type_known
+# endif
+# if defined(LINUX) && defined(__x86_64__)
+#    define X86_64
+#    define mach_type_known
+# endif
+# if defined(LINUX) && (defined(__ia64__) || defined(__ia64))
+#    define IA64
+#    define mach_type_known
+# endif
+# if defined(LINUX) && defined(__arm__)
+#    define ARM32
+#    define mach_type_known
+# endif
+# if defined(LINUX) && defined(__cris__)
+#    ifndef CRIS
+#	define CRIS
+#    endif
+#    define mach_type_known
+# endif
+# if defined(LINUX) && (defined(powerpc) || defined(__powerpc__) || \
+		        defined(powerpc64) || defined(__powerpc64__))
+#    define POWERPC
+#    define mach_type_known
+# endif
+# if defined(FREEBSD) && (defined(powerpc) || defined(__powerpc__))
+#    define POWERPC
+#    define mach_type_known
+# endif
+# if defined(LINUX) && defined(__mc68000__)
+#    define M68K
+#    define mach_type_known
+# endif
+# if defined(LINUX) && (defined(sparc) || defined(__sparc__))
+#    define SPARC
+#    define mach_type_known
+# endif
+# if defined(LINUX) && defined(__arm__)
+#    define ARM32
+#    define mach_type_known
+# endif
+# if defined(LINUX) && defined(__sh__)
+#    define SH
+#    define mach_type_known
+# endif
+# if defined(LINUX) && defined(__m32r__)
+#    define M32R
+#    define mach_type_known
+# endif
+# if defined(__alpha) || defined(__alpha__)
+#   define ALPHA
+#   if !defined(LINUX) && !defined(NETBSD) && !defined(OPENBSD) && !defined(FREEBSD)
+#     define OSF1	/* a.k.a Digital Unix */
+#   endif
+#   define mach_type_known
+# endif
+# if defined(_AMIGA) && !defined(AMIGA)
+#   define AMIGA
+# endif
+# ifdef AMIGA 
+#   define M68K
+#   define mach_type_known
+# endif
+# if defined(THINK_C) || defined(__MWERKS__) && !defined(__powerc)
+#   define M68K
+#   define MACOS
+#   define mach_type_known
+# endif
+# if defined(__MWERKS__) && defined(__powerc) && !defined(__MACH__)
+#   define POWERPC
+#   define MACOS
+#   define mach_type_known
+# endif
+# if defined(macosx) || (defined(__APPLE__) && defined(__MACH__))
+#   define DARWIN
+#   if defined(__ppc__)  || defined(__ppc64__)
+#    define POWERPC
+#    define mach_type_known
+#   elif defined(__x86_64__)
+#    define X86_64
+#    define mach_type_known
+#   elif defined(__i386__)
+#    define I386
+#    define mach_type_known
+#   endif
+# endif
+# if defined(NeXT) && defined(mc68000)
+#   define M68K
+#   define NEXT
+#   define mach_type_known
+# endif
+# if defined(NeXT) && (defined(i386) || defined(__i386__))
+#   define I386
+#   define NEXT
+#   define mach_type_known
+# endif
+# if defined(__OpenBSD__) && (defined(i386) || defined(__i386__))
+#   define I386
+#   define OPENBSD
+#   define mach_type_known
+# endif
+# if defined(FREEBSD) && (defined(i386) || defined(__i386__))
+#   define I386
+#   define mach_type_known
+# endif
+# if defined(FREEBSD) && defined(__x86_64__)
+#   define X86_64
+#   define mach_type_known
+# endif
+# if defined(__NetBSD__) && (defined(i386) || defined(__i386__))
+#   define I386
+#   define mach_type_known
+# endif
+# if defined(__NetBSD__) && defined(__x86_64__)
+#    define X86_64
+#    define mach_type_known
+# endif
+# if defined(FREEBSD) && defined(__sparc__)
+#    define SPARC
+#    define mach_type_known
+#endif
+# if defined(bsdi) && (defined(i386) || defined(__i386__))
+#    define I386
+#    define BSDI
+#    define mach_type_known
+# endif
+# if !defined(mach_type_known) && defined(__386BSD__)
+#   define I386
+#   define THREE86BSD
+#   define mach_type_known
+# endif
+# if defined(_CX_UX) && defined(_M88K)
+#   define M88K
+#   define CX_UX
+#   define mach_type_known
+# endif
+# if defined(DGUX) && defined(m88k)
+#   define M88K
+    /* DGUX defined */
+#   define mach_type_known
+# endif
+# if defined(_WIN32_WCE)
+    /* SH3, SH4, MIPS already defined for corresponding architectures */
+#   if defined(SH3) || defined(SH4)
+#     define SH
+#   endif
+#   if defined(x86)
+#     define I386
+#   endif
+#   if defined(ARM)
+#     define ARM32
+#   endif
+#   define MSWINCE
+#   define mach_type_known
+# else
+#   if (defined(_MSDOS) || defined(_MSC_VER)) && (_M_IX86 >= 300) \
+        || defined(_WIN32) && !defined(__CYGWIN32__) && !defined(__CYGWIN__)
+#     define I386
+#     define MSWIN32	/* or Win32s */
+#     define mach_type_known
+#   endif
+#   if defined(_MSC_VER) && defined(_M_IA64)
+#     define IA64
+#     define MSWIN32	/* Really win64, but we don't treat 64-bit 	*/
+			/* variants as a differnt platform.		*/
+#   endif
+# endif
+# if defined(__DJGPP__)
+#   define I386
+#   ifndef DJGPP
+#     define DJGPP  /* MSDOS running the DJGPP port of GCC */
+#   endif
+#   define mach_type_known
+# endif
+# if defined(__CYGWIN32__) || defined(__CYGWIN__)
+#   define I386
+#   define CYGWIN32
+#   define mach_type_known
+# endif
+# if defined(__MINGW32__)
+#   define I386
+#   define MSWIN32
+#   define mach_type_known
+# endif
+# if defined(__BORLANDC__)
+#   define I386
+#   define MSWIN32
+#   define mach_type_known
+# endif
+# if defined(_UTS) && !defined(mach_type_known)
+#   define S370
+#   define UTS4
+#   define mach_type_known
+# endif
+# if defined(__pj__)
+#   define PJ
+#   define mach_type_known
+# endif
+# if defined(__embedded__) && defined(PPC)
+#   define POWERPC
+#   define NOSYS
+#   define mach_type_known
+# endif
+/* Ivan Demakov */
+# if defined(__WATCOMC__) && defined(__386__)
+#   define I386
+#   if !defined(OS2) && !defined(MSWIN32) && !defined(DOS4GW)
+#     if defined(__OS2__)
+#       define OS2
+#     else
+#       if defined(__WINDOWS_386__) || defined(__NT__)
+#         define MSWIN32
+#       else
+#         define DOS4GW
+#       endif
+#     endif
+#   endif
+#   define mach_type_known
+# endif
+# if defined(__s390__) && defined(LINUX)
+#    define S390
+#    define mach_type_known
+# endif
+# if defined(__GNU__)
+#   if defined(__i386__)
+/* The Debian Hurd running on generic PC */  
+#     define  HURD
+#     define  I386
+#     define  mach_type_known
+#    endif 
+# endif
+
+/* Feel free to add more clauses here */
+
+/* Or manually define the machine type here.  A machine type is 	*/
+/* characterized by the architecture.  Some				*/
+/* machine types are further subdivided by OS.				*/
+/* the macros ULTRIX, RISCOS, and BSD to distinguish.			*/
+/* Note that SGI IRIX is treated identically to RISCOS.			*/
+/* SYSV on an M68K actually means A/UX.					*/
+/* The distinction in these cases is usually the stack starting address */
+# ifndef mach_type_known
+	--> unknown machine type
+# endif
+		    /* Mapping is: M68K       ==> Motorola 680X0	*/
+		    /*		   (SUNOS4,HP,NEXT, and SYSV (A/UX),	*/
+		    /*		   MACOS and AMIGA variants)		*/
+		    /*             I386       ==> Intel 386	 	*/
+		    /*		    (SEQUENT, OS2, SCO, LINUX, NETBSD,	*/
+		    /*		     FREEBSD, THREE86BSD, MSWIN32,	*/
+		    /* 		     BSDI,SUNOS5, NEXT, other variants)	*/
+                    /*             NS32K      ==> Encore Multimax 	*/
+                    /*             MIPS       ==> R2000 or R3000	*/
+                    /*			(RISCOS, ULTRIX variants)	*/
+                    /*		   VAX	      ==> DEC VAX		*/
+                    /*			(BSD, ULTRIX variants)		*/
+                    /*		   RS6000     ==> IBM RS/6000 AIX3.X	*/
+                    /*		   RT	      ==> IBM PC/RT		*/
+                    /*		   HP_PA      ==> HP9000/700 & /800	*/
+                    /*				  HP/UX, LINUX		*/
+		    /*		   SPARC      ==> SPARC	v7/v8/v9	*/
+		    /*			(SUNOS4, SUNOS5, LINUX,		*/
+		    /*			 DRSNX variants)		*/
+		    /* 		   ALPHA      ==> DEC Alpha 		*/
+		    /*			(OSF1 and LINUX variants)	*/
+		    /* 		   M88K       ==> Motorola 88XX0        */
+		    /* 		        (CX_UX and DGUX)		*/
+		    /* 		   S370	      ==> 370-like machine	*/
+		    /* 			running Amdahl UTS4		*/
+                    /*             S390       ==> 390-like machine      */
+		    /*                  running LINUX                   */
+		    /* 		   ARM32      ==> Intel StrongARM	*/
+		    /* 		   IA64	      ==> Intel IPF		*/
+		    /*				  (e.g. Itanium)	*/
+		    /*			(LINUX and HPUX)	        */
+		    /*		   SH	      ==> Hitachi SuperH	*/
+		    /* 			(LINUX & MSWINCE)		*/
+		    /* 		   X86_64     ==> AMD x86-64		*/
+		    /*		   POWERPC    ==> IBM/Apple PowerPC	*/
+		    /*			(MACOS(<=9),DARWIN(incl.MACOSX),*/
+		    /*			 LINUX, NETBSD, NOSYS variants)	*/
+		    /*			Handles 32 and 64-bit variants.	*/
+		    /* 			AIX should be handled here, but	*/
+		    /*			that's called an RS6000.	*/
+		    /*		   CRIS       ==> Axis Etrax		*/
+		    /*		   M32R	      ==> Renesas M32R		*/
+
+
+/*
+ * For each architecture and OS, the following need to be defined:
+ *
+ * CPP_WORDSZ is a simple integer constant representing the word size.
+ * in bits.  We assume byte addressibility, where a byte has 8 bits.
+ * We also assume CPP_WORDSZ is either 32 or 64.
+ * (We care about the length of pointers, not hardware
+ * bus widths.  Thus a 64 bit processor with a C compiler that uses
+ * 32 bit pointers should use CPP_WORDSZ of 32, not 64. Default is 32.)
+ *
+ * MACH_TYPE is a string representation of the machine type.
+ * OS_TYPE is analogous for the OS.
+ *
+ * ALIGNMENT is the largest N, such that
+ * all pointer are guaranteed to be aligned on N byte boundaries.
+ * defining it to be 1 will always work, but perform poorly.
+ *
+ * DATASTART is the beginning of the data segment.
+ * On some platforms SEARCH_FOR_DATA_START is defined.
+ * SEARCH_FOR_DATASTART will cause GC_data_start to
+ * be set to an address determined by accessing data backwards from _end
+ * until an unmapped page is found.  DATASTART will be defined to be
+ * GC_data_start.
+ * On UNIX-like systems, the collector will scan the area between DATASTART
+ * and DATAEND for root pointers.
+ *
+ * DATAEND, if not `end' where `end' is defined as ``extern int end[];''.
+ * RTH suggests gaining access to linker script synth'd values with
+ * this idiom instead of `&end' where `end' is defined as ``extern int end;'' .
+ * Otherwise, ``GCC will assume these are in .sdata/.sbss'' and it will, e.g.,
+ * cause failures on alpha*-*-* with ``-msmall-data or -fpic'' or mips-*-*
+ * without any special options.
+ *
+ * ALIGN_DOUBLE of GC_malloc should return blocks aligned to twice
+ * the pointer size.
+ *
+ * STACKBOTTOM is the cool end of the stack, which is usually the
+ * highest address in the stack.
+ * Under PCR or OS/2, we have other ways of finding thread stacks.
+ * For each machine, the following should:
+ * 1) define STACK_GROWS_UP if the stack grows toward higher addresses, and
+ * 2) define exactly one of
+ *	STACKBOTTOM (should be defined to be an expression)
+ *	LINUX_STACKBOTTOM
+ *	HEURISTIC1
+ *	HEURISTIC2
+ * If STACKBOTTOM is defined, then it's value will be used directly as the
+ * stack base.  If LINUX_STACKBOTTOM is defined, then it will be determined
+ * with a method appropriate for most Linux systems.  Currently we look
+ * first for __libc_stack_end, and if that fails read it from /proc.
+ * If either of the last two macros are defined, then STACKBOTTOM is computed
+ * during collector startup using one of the following two heuristics:
+ * HEURISTIC1:  Take an address inside GC_init's frame, and round it up to
+ *		the next multiple of STACK_GRAN.
+ * HEURISTIC2:  Take an address inside GC_init's frame, increment it repeatedly
+ *		in small steps (decrement if STACK_GROWS_UP), and read the value
+ *		at each location.  Remember the value when the first
+ *		Segmentation violation or Bus error is signalled.  Round that
+ *		to the nearest plausible page boundary, and use that instead
+ *		of STACKBOTTOM.
+ *
+ * Gustavo Rodriguez-Rivera points out that on most (all?) Unix machines,
+ * the value of environ is a pointer that can serve as STACKBOTTOM.
+ * I expect that HEURISTIC2 can be replaced by this approach, which
+ * interferes far less with debugging.  However it has the disadvantage
+ * that it's confused by a putenv call before the collector is initialized.
+ * This could be dealt with by intercepting putenv ...
+ *
+ * If no expression for STACKBOTTOM can be found, and neither of the above
+ * heuristics are usable, the collector can still be used with all of the above
+ * undefined, provided one of the following is done:
+ * 1) GC_mark_roots can be changed to somehow mark from the correct stack(s)
+ *    without reference to STACKBOTTOM.  This is appropriate for use in
+ *    conjunction with thread packages, since there will be multiple stacks.
+ *    (Allocating thread stacks in the heap, and treating them as ordinary
+ *    heap data objects is also possible as a last resort.  However, this is
+ *    likely to introduce significant amounts of excess storage retention
+ *    unless the dead parts of the thread stacks are periodically cleared.)
+ * 2) Client code may set GC_stackbottom before calling any GC_ routines.
+ *    If the author of the client code controls the main program, this is
+ *    easily accomplished by introducing a new main program, setting
+ *    GC_stackbottom to the address of a local variable, and then calling
+ *    the original main program.  The new main program would read something
+ *    like:
+ *
+ *		# include "gc_private.h"
+ *
+ *		main(argc, argv, envp)
+ *		int argc;
+ *		char **argv, **envp;
+ *		{
+ *		    int dummy;
+ *
+ *		    GC_stackbottom = (ptr_t)(&dummy);
+ *		    return(real_main(argc, argv, envp));
+ *		}
+ *
+ *
+ * Each architecture may also define the style of virtual dirty bit
+ * implementation to be used:
+ *   MPROTECT_VDB: Write protect the heap and catch faults.
+ *   PROC_VDB: Use the SVR4 /proc primitives to read dirty bits.
+ *
+ * An architecture may define DYNAMIC_LOADING if dynamic_load.c
+ * defined GC_register_dynamic_libraries() for the architecture.
+ *
+ * An architecture may define PREFETCH(x) to preload the cache with *x.
+ * This defaults to a no-op.
+ *
+ * PREFETCH_FOR_WRITE(x) is used if *x is about to be written.
+ *
+ * An architecture may also define CLEAR_DOUBLE(x) to be a fast way to
+ * clear the two words at GC_malloc-aligned address x.  By default,
+ * word stores of 0 are used instead.
+ *
+ * HEAP_START may be defined as the initial address hint for mmap-based
+ * allocation.
+ */
+
+/* If we are using a recent version of gcc, we can use __builtin_unwind_init()
+ * to push the relevant registers onto the stack.  This generally makes
+ * USE_GENERIC_PUSH_REGS the preferred approach for marking from registers.
+ */
+# if defined(__GNUC__) && ((__GNUC__ >= 3) || \
+			   (__GNUC__ == 2 && __GNUC_MINOR__ >= 8)) \
+		       && !defined(__INTEL_COMPILER) \
+ 		       && !defined(__PATHCC__)
+#   define HAVE_BUILTIN_UNWIND_INIT
+# endif
+
+# define STACK_GRAN 0x1000000
+# ifdef M68K
+#   define MACH_TYPE "M68K"
+#   define ALIGNMENT 2
+#   ifdef OPENBSD
+#	define OS_TYPE "OPENBSD"
+#	define HEURISTIC2
+#	ifdef __ELF__
+#	  define DATASTART GC_data_start
+#	  define DYNAMIC_LOADING
+#	else
+	  extern char etext[];
+#	  define DATASTART ((ptr_t)(etext))
+#       endif
+#       define USE_GENERIC_PUSH_REGS
+#   endif
+#   ifdef NETBSD
+#	define OS_TYPE "NETBSD"
+#	define HEURISTIC2
+#	ifdef __ELF__
+#	  define DATASTART GC_data_start
+#	  define DYNAMIC_LOADING
+#	else
+	  extern char etext[];
+#	  define DATASTART ((ptr_t)(etext))
+#       endif
+#	define USE_GENERIC_PUSH_REGS
+#   endif
+#   ifdef LINUX
+#       define OS_TYPE "LINUX"
+#       define LINUX_STACKBOTTOM
+#       define USE_GENERIC_PUSH_REGS
+		/* We never got around to the assembly version. */
+#       define MPROTECT_VDB
+#       ifdef __ELF__
+#            define DYNAMIC_LOADING
+#	     include <features.h>
+#	     if defined(__GLIBC__)&& __GLIBC__>=2
+#              define SEARCH_FOR_DATA_START
+#	     else /* !GLIBC2 */
+               extern char **__environ;
+#              define DATASTART ((ptr_t)(&__environ))
+                             /* hideous kludge: __environ is the first */
+                             /* word in crt0.o, and delimits the start */
+                             /* of the data segment, no matter which   */
+                             /* ld options were passed through.        */
+                             /* We could use _etext instead, but that  */
+                             /* would include .rodata, which may       */
+                             /* contain large read-only data tables    */
+                             /* that we'd rather not scan.             */
+#	     endif /* !GLIBC2 */
+             extern int _end[];
+#            define DATAEND (_end)
+#       else
+             extern int etext[];
+#            define DATASTART ((ptr_t)((((word) (etext)) + 0xfff) & ~0xfff))
+#       endif
+#   endif
+#   ifdef SUNOS4
+#	define OS_TYPE "SUNOS4"
+	extern char etext[];
+#	define DATASTART ((ptr_t)((((word) (etext)) + 0x1ffff) & ~0x1ffff))
+#	define HEURISTIC1	/* differs	*/
+#	define DYNAMIC_LOADING
+#   endif
+#   ifdef HP
+#	define OS_TYPE "HP"
+	extern char etext[];
+#       define DATASTART ((ptr_t)((((word) (etext)) + 0xfff) & ~0xfff))
+#       define STACKBOTTOM ((ptr_t) 0xffeffffc)
+			      /* empirically determined.  seems to work. */
+#  	include <unistd.h>
+#	define GETPAGESIZE() sysconf(_SC_PAGE_SIZE)
+#   endif
+#   ifdef SYSV
+#	define OS_TYPE "SYSV"
+	extern etext[];
+#   	define DATASTART ((ptr_t)((((word) (etext)) + 0x3fffff) \
+				   & ~0x3fffff) \
+				  +((word)etext & 0x1fff))
+	/* This only works for shared-text binaries with magic number 0413.
+	   The other sorts of SysV binaries put the data at the end of the text,
+	   in which case the default of etext would work.  Unfortunately,
+	   handling both would require having the magic-number available.
+	   	   		-- Parag
+	   */
+#	define STACKBOTTOM ((ptr_t)0xFFFFFFFE)
+			/* The stack starts at the top of memory, but   */
+			/* 0x0 cannot be used as setjump_test complains */
+			/* that the stack direction is incorrect.  Two  */
+			/* bytes down from 0x0 should be safe enough.   */
+			/* 		--Parag				*/
+#   	include <sys/mmu.h>
+#	define GETPAGESIZE() PAGESIZE	/* Is this still right? */
+#   endif
+#   ifdef AMIGA
+#	define OS_TYPE "AMIGA"
+ 	    	/* STACKBOTTOM and DATASTART handled specially	*/
+ 	    	/* in os_dep.c					*/
+# 	define DATAEND	/* not needed */
+#	define GETPAGESIZE() 4096
+#   endif
+#   ifdef MACOS
+#     ifndef __LOWMEM__
+#     include <LowMem.h>
+#     endif
+#     define OS_TYPE "MACOS"
+			/* see os_dep.c for details of global data segments. */
+#     define STACKBOTTOM ((ptr_t) LMGetCurStackBase())
+#     define DATAEND	/* not needed */
+#     define GETPAGESIZE() 4096
+#   endif
+#   ifdef NEXT
+#	define OS_TYPE "NEXT"
+#	define DATASTART ((ptr_t) get_etext())
+#	define STACKBOTTOM ((ptr_t) 0x4000000)
+#	define DATAEND	/* not needed */
+#   endif
+# endif
+
+# if defined(POWERPC)
+#   define MACH_TYPE "POWERPC"
+#   ifdef MACOS
+#     define ALIGNMENT 2  /* Still necessary?  Could it be 4?	*/
+#     ifndef __LOWMEM__
+#     include <LowMem.h>
+#     endif
+#     define OS_TYPE "MACOS"
+			/* see os_dep.c for details of global data segments. */
+#     define STACKBOTTOM ((ptr_t) LMGetCurStackBase())
+#     define DATAEND  /* not needed */
+#   endif
+#   ifdef LINUX
+#     if defined(__powerpc64__)
+#       define ALIGNMENT 8
+#       define CPP_WORDSZ 64
+#       ifndef HBLKSIZE
+#         define HBLKSIZE 4096
+#       endif
+#     else
+#       define ALIGNMENT 4
+#     endif
+#     define OS_TYPE "LINUX"
+      /* HEURISTIC1 has been reliably reported to fail for a 32-bit	*/
+      /* executable on a 64 bit kernel.					*/
+#     define LINUX_STACKBOTTOM
+#     define DYNAMIC_LOADING
+#     define SEARCH_FOR_DATA_START
+      extern int _end[];
+#     define DATAEND (_end)
+#   endif
+#   ifdef DARWIN
+#     define OS_TYPE "DARWIN"
+#     define DYNAMIC_LOADING
+#     if defined(__ppc64__)
+#       define ALIGNMENT 8
+#       define CPP_WORDSZ 64
+#       define STACKBOTTOM ((ptr_t) 0x7fff5fc00000)
+#       define CACHE_LINE_SIZE 64
+#       ifndef HBLKSIZE
+#         define HBLKSIZE 4096
+#       endif
+#     else
+#       define ALIGNMENT 4
+#       define STACKBOTTOM ((ptr_t) 0xc0000000)
+#     endif
+      /* XXX: see get_end(3), get_etext() and get_end() should not be used.
+	 These aren't used when dyld support is enabled (it is by default) */
+#     define DATASTART ((ptr_t) get_etext())
+#     define DATAEND	((ptr_t) get_end())
+#     define USE_MMAP
+#     define USE_MMAP_ANON
+#     define USE_ASM_PUSH_REGS
+#     ifdef GC_DARWIN_THREADS
+#       define MPROTECT_VDB
+#     endif
+#     include <unistd.h>
+#     define GETPAGESIZE() getpagesize()
+#     if defined(USE_PPC_PREFETCH) && defined(__GNUC__)
+	/* The performance impact of prefetches is untested */
+#	define PREFETCH(x) \
+	  __asm__ __volatile__ ("dcbt 0,%0" : : "r" ((const void *) (x)))
+#	define PREFETCH_FOR_WRITE(x) \
+	  __asm__ __volatile__ ("dcbtst 0,%0" : : "r" ((const void *) (x)))
+#     endif
+      /* There seems to be some issues with trylock hanging on darwin. This
+	 should be looked into some more */
+#     define NO_PTHREAD_TRYLOCK
+#   endif
+#   ifdef FREEBSD
+#       define ALIGNMENT 4
+#       define OS_TYPE "FREEBSD"
+#       ifndef GC_FREEBSD_THREADS
+#           define MPROTECT_VDB
+#       endif
+#       define SIG_SUSPEND SIGUSR1
+#       define SIG_THR_RESTART SIGUSR2
+#       define FREEBSD_STACKBOTTOM
+#       ifdef __ELF__
+#           define DYNAMIC_LOADING
+#       endif
+        extern char etext[];
+        extern char * GC_FreeBSDGetDataStart();
+#       define DATASTART GC_FreeBSDGetDataStart(0x1000, &etext)
+#   endif
+#   ifdef NETBSD
+#     define ALIGNMENT 4
+#     define OS_TYPE "NETBSD"
+#     define HEURISTIC2
+      extern char etext[];
+#     define DATASTART GC_data_start
+#     define DYNAMIC_LOADING
+#   endif
+#   ifdef NOSYS
+#     define ALIGNMENT 4
+#     define OS_TYPE "NOSYS"
+      extern void __end[], __dso_handle[];
+#     define DATASTART (__dso_handle)  /* OK, that's ugly.  */
+#     define DATAEND (__end)
+	/* Stack starts at 0xE0000000 for the simulator.  */
+#     undef STACK_GRAN
+#     define STACK_GRAN 0x10000000
+#     define HEURISTIC1
+#   endif
+# endif
+
+# ifdef VAX
+#   define MACH_TYPE "VAX"
+#   define ALIGNMENT 4	/* Pointers are longword aligned by 4.2 C compiler */
+    extern char etext[];
+#   define DATASTART ((ptr_t)(etext))
+#   ifdef BSD
+#	define OS_TYPE "BSD"
+#	define HEURISTIC1
+			/* HEURISTIC2 may be OK, but it's hard to test. */
+#   endif
+#   ifdef ULTRIX
+#	define OS_TYPE "ULTRIX"
+#	define STACKBOTTOM ((ptr_t) 0x7fffc800)
+#   endif
+# endif
+
+# ifdef RT
+#   define MACH_TYPE "RT"
+#   define ALIGNMENT 4
+#   define DATASTART ((ptr_t) 0x10000000)
+#   define STACKBOTTOM ((ptr_t) 0x1fffd800)
+# endif
+
+# ifdef SPARC
+#   define MACH_TYPE "SPARC"
+#   if defined(__arch64__) || defined(__sparcv9)
+#     define ALIGNMENT 8
+#     define CPP_WORDSZ 64
+#     define ELF_CLASS ELFCLASS64
+#   else
+#     define ALIGNMENT 4	/* Required by hardware	*/
+#     define CPP_WORDSZ 32
+#   endif
+#   define ALIGN_DOUBLE
+#   ifdef SUNOS5
+#	define OS_TYPE "SUNOS5"
+	extern int _etext[];
+	extern int _end[];
+	extern ptr_t GC_SysVGetDataStart();
+#       define DATASTART GC_SysVGetDataStart(0x10000, _etext)
+#	define DATAEND (_end)
+#	if !defined(USE_MMAP) && defined(REDIRECT_MALLOC)
+#	    define USE_MMAP
+	    /* Otherwise we now use calloc.  Mmap may result in the	*/
+	    /* heap interleaved with thread stacks, which can result in	*/
+	    /* excessive blacklisting.  Sbrk is unusable since it	*/
+	    /* doesn't interact correctly with the system malloc.	*/
+#	endif
+#       ifdef USE_MMAP
+#         define HEAP_START (ptr_t)0x40000000
+#       else
+#	  define HEAP_START DATAEND
+#       endif
+#	define PROC_VDB
+/*	HEURISTIC1 reportedly no longer works under 2.7.  		*/
+/*  	HEURISTIC2 probably works, but this appears to be preferable.	*/
+/*	Apparently USRSTACK is defined to be USERLIMIT, but in some	*/
+/* 	installations that's undefined.  We work around this with a	*/
+/*	gross hack:							*/
+#       include <sys/vmparam.h>
+#	ifdef USERLIMIT
+	  /* This should work everywhere, but doesn't.	*/
+#	  define STACKBOTTOM USRSTACK
+#       else
+#	  define HEURISTIC2
+#       endif
+#	include <unistd.h>
+#       define GETPAGESIZE()  sysconf(_SC_PAGESIZE)
+		/* getpagesize() appeared to be missing from at least one */
+		/* Solaris 5.4 installation.  Weird.			  */
+#	define DYNAMIC_LOADING
+#   endif
+#   ifdef SUNOS4
+#	define OS_TYPE "SUNOS4"
+	/* [If you have a weak stomach, don't read this.]		*/
+	/* We would like to use:					*/
+/* #       define DATASTART ((ptr_t)((((word) (etext)) + 0x1fff) & ~0x1fff)) */
+	/* This fails occasionally, due to an ancient, but very 	*/
+	/* persistent ld bug.  etext is set 32 bytes too high.		*/
+	/* We instead read the text segment size from the a.out		*/
+	/* header, which happens to be mapped into our address space	*/
+	/* at the start of the text segment.  The detective work here	*/
+	/* was done by Robert Ehrlich, Manuel Serrano, and Bernard	*/
+	/* Serpette of INRIA.						*/
+	/* This assumes ZMAGIC, i.e. demand-loadable executables.	*/
+#	define TEXTSTART 0x2000
+#       define DATASTART ((ptr_t)(*(int *)(TEXTSTART+0x4)+TEXTSTART))
+#	define MPROTECT_VDB
+#	define HEURISTIC1
+# 	define DYNAMIC_LOADING
+#   endif
+#   ifdef DRSNX
+#	define OS_TYPE "DRSNX"
+	extern ptr_t GC_SysVGetDataStart();
+	extern int etext[];
+#       define DATASTART GC_SysVGetDataStart(0x10000, etext)
+#	define MPROTECT_VDB
+#       define STACKBOTTOM ((ptr_t) 0xdfff0000)
+#	define DYNAMIC_LOADING
+#   endif
+#   ifdef LINUX
+#     define OS_TYPE "LINUX"
+#     ifdef __ELF__
+#       define DYNAMIC_LOADING
+#     else
+          Linux Sparc/a.out not supported
+#     endif
+      extern int _end[];
+      extern int _etext[];
+#     define DATAEND (_end)
+#     define SVR4
+      extern ptr_t GC_SysVGetDataStart();
+#     ifdef __arch64__
+#	define DATASTART GC_SysVGetDataStart(0x100000, _etext)
+#     else
+#       define DATASTART GC_SysVGetDataStart(0x10000, _etext)
+#     endif
+#     define LINUX_STACKBOTTOM
+#   endif
+#   ifdef OPENBSD
+#     define OS_TYPE "OPENBSD"
+#     define STACKBOTTOM ((ptr_t) 0xf8000000)
+      extern int etext[];
+#     define DATASTART ((ptr_t)(etext))
+#   endif
+#   ifdef NETBSD
+#     define OS_TYPE "NETBSD"
+#     define HEURISTIC2
+#     ifdef __ELF__
+#	define DATASTART GC_data_start
+#	define DYNAMIC_LOADING
+#     else
+	extern char etext[];
+#	define DATASTART ((ptr_t)(etext))
+#     endif
+#   endif
+#   ifdef FREEBSD
+#	define OS_TYPE "FREEBSD"
+#	define SIG_SUSPEND SIGUSR1
+#	define SIG_THR_RESTART SIGUSR2
+#	define FREEBSD_STACKBOTTOM
+#	ifdef __ELF__
+#	    define DYNAMIC_LOADING
+#	endif
+	extern char etext[];
+	extern char edata[];
+	extern char end[];
+#	define NEED_FIND_LIMIT
+#	define DATASTART ((ptr_t)(&etext))
+#	define DATAEND (GC_find_limit (DATASTART, TRUE))
+#	define DATASTART2 ((ptr_t)(&edata))
+#	define DATAEND2 ((ptr_t)(&end))
+#   endif
+# endif
+
+# ifdef I386
+#   define MACH_TYPE "I386"
+#   if defined(__LP64__) || defined(_WIN64)
+#     define CPP_WORDSZ 64
+#     define ALIGNMENT 8
+#   else
+#     define CPP_WORDSZ 32
+#     define ALIGNMENT 4
+			/* Appears to hold for all "32 bit" compilers	*/
+			/* except Borland.  The -a4 option fixes 	*/
+			/* Borland.					*/
+                        /* Ivan Demakov: For Watcom the option is -zp4. */
+#   endif
+#   ifndef SMALL_CONFIG
+#     define ALIGN_DOUBLE /* Not strictly necessary, but may give speed   */
+			  /* improvement on Pentiums.			  */
+#   endif
+#   ifdef HAVE_BUILTIN_UNWIND_INIT
+#	define USE_GENERIC_PUSH_REGS
+#   endif
+#   ifdef SEQUENT
+#	define OS_TYPE "SEQUENT"
+	extern int etext[];
+#       define DATASTART ((ptr_t)((((word) (etext)) + 0xfff) & ~0xfff))
+#       define STACKBOTTOM ((ptr_t) 0x3ffff000) 
+#   endif
+#   ifdef BEOS
+#     define OS_TYPE "BEOS"
+#     include <OS.h>
+#     define GETPAGESIZE() B_PAGE_SIZE
+      extern int etext[];
+#     define DATASTART ((ptr_t)((((word) (etext)) + 0xfff) & ~0xfff))
+#   endif
+#   ifdef SUNOS5
+#	define OS_TYPE "SUNOS5"
+        extern int _etext[], _end[];
+  	extern ptr_t GC_SysVGetDataStart();
+#       define DATASTART GC_SysVGetDataStart(0x1000, _etext)
+#	define DATAEND (_end)
+/*	# define STACKBOTTOM ((ptr_t)(_start)) worked through 2.7,  	*/
+/*      but reportedly breaks under 2.8.  It appears that the stack	*/
+/* 	base is a property of the executable, so this should not break	*/
+/* 	old executables.						*/
+/*  	HEURISTIC2 probably works, but this appears to be preferable.	*/
+#       include <sys/vm.h>
+#	define STACKBOTTOM USRSTACK
+/* At least in Solaris 2.5, PROC_VDB gives wrong values for dirty bits. */
+/* It appears to be fixed in 2.8 and 2.9.				*/
+#	ifdef SOLARIS25_PROC_VDB_BUG_FIXED
+#	  define PROC_VDB
+#	endif
+#	define DYNAMIC_LOADING
+#	if !defined(USE_MMAP) && defined(REDIRECT_MALLOC)
+#	    define USE_MMAP
+	    /* Otherwise we now use calloc.  Mmap may result in the	*/
+	    /* heap interleaved with thread stacks, which can result in	*/
+	    /* excessive blacklisting.  Sbrk is unusable since it	*/
+	    /* doesn't interact correctly with the system malloc.	*/
+#	endif
+#       ifdef USE_MMAP
+#         define HEAP_START (ptr_t)0x40000000
+#       else
+#	  define HEAP_START DATAEND
+#       endif
+#   endif
+#   ifdef SCO
+#	define OS_TYPE "SCO"
+	extern int etext[];
+#   	define DATASTART ((ptr_t)((((word) (etext)) + 0x3fffff) \
+				  & ~0x3fffff) \
+				 +((word)etext & 0xfff))
+#	define STACKBOTTOM ((ptr_t) 0x7ffffffc)
+#   endif
+#   ifdef SCO_ELF
+#       define OS_TYPE "SCO_ELF"
+        extern int etext[];
+#       define DATASTART ((ptr_t)(etext))
+#       define STACKBOTTOM ((ptr_t) 0x08048000)
+#       define DYNAMIC_LOADING
+#	define ELF_CLASS ELFCLASS32
+#   endif
+#   ifdef DGUX
+#	define OS_TYPE "DGUX"
+	extern int _etext, _end;
+	extern ptr_t GC_SysVGetDataStart();
+#	define DATASTART GC_SysVGetDataStart(0x1000, &_etext)
+#	define DATAEND (&_end)
+#	define STACK_GROWS_DOWN
+#	define HEURISTIC2
+#	include <unistd.h>
+#	define GETPAGESIZE()  sysconf(_SC_PAGESIZE)
+#	define DYNAMIC_LOADING
+#	ifndef USE_MMAP
+#	  define USE_MMAP
+#	endif /* USE_MMAP */
+#	define MAP_FAILED (void *) -1
+#	ifdef USE_MMAP
+#	  define HEAP_START (ptr_t)0x40000000
+#	else /* USE_MMAP */
+#	  define HEAP_START DATAEND
+#	endif /* USE_MMAP */
+#   endif /* DGUX */
+
+#   ifdef LINUX
+#	ifndef __GNUC__
+	  /* The Intel compiler doesn't like inline assembly */
+#	  define USE_GENERIC_PUSH_REGS
+# 	endif
+#	define OS_TYPE "LINUX"
+#       define LINUX_STACKBOTTOM
+#	if 0
+#	  define HEURISTIC1
+#         undef STACK_GRAN
+#         define STACK_GRAN 0x10000000
+	  /* STACKBOTTOM is usually 0xc0000000, but this changes with	*/
+	  /* different kernel configurations.  In particular, systems	*/
+	  /* with 2GB physical memory will usually move the user	*/
+	  /* address space limit, and hence initial SP to 0x80000000.	*/
+#       endif
+#       if !defined(GC_LINUX_THREADS) || !defined(REDIRECT_MALLOC)
+#	    define MPROTECT_VDB
+#	else
+	    /* We seem to get random errors in incremental mode,	*/
+	    /* possibly because Linux threads is itself a malloc client */
+	    /* and can't deal with the signals.				*/
+#	endif
+#	define HEAP_START (ptr_t)0x1000
+		/* This encourages mmap to give us low addresses,	*/
+		/* thus allowing the heap to grow to ~3GB		*/
+#       ifdef __ELF__
+#            define DYNAMIC_LOADING
+#	     ifdef UNDEFINED	/* includes ro data */
+	       extern int _etext[];
+#              define DATASTART ((ptr_t)((((word) (_etext)) + 0xfff) & ~0xfff))
+#	     endif
+#	     include <features.h>
+#	     if defined(__GLIBC__) && __GLIBC__ >= 2
+#		 define SEARCH_FOR_DATA_START
+#	     else
+     	         extern char **__environ;
+#                define DATASTART ((ptr_t)(&__environ))
+			      /* hideous kludge: __environ is the first */
+			      /* word in crt0.o, and delimits the start */
+			      /* of the data segment, no matter which   */
+			      /* ld options were passed through.        */
+			      /* We could use _etext instead, but that  */
+			      /* would include .rodata, which may       */
+			      /* contain large read-only data tables    */
+			      /* that we'd rather not scan.		*/
+#	     endif
+	     extern int _end[];
+#	     define DATAEND (_end)
+#	else
+	     extern int etext[];
+#            define DATASTART ((ptr_t)((((word) (etext)) + 0xfff) & ~0xfff))
+#       endif
+#	ifdef USE_I686_PREFETCH
+	  /* FIXME: Thus should use __builtin_prefetch, but we'll leave that	*/
+	  /* for the next rtelease.						*/
+#	  define PREFETCH(x) \
+	    __asm__ __volatile__ ("	prefetchnta	%0": : "m"(*(char *)(x)))
+	    /* Empirically prefetcht0 is much more effective at reducing	*/
+	    /* cache miss stalls for the targetted load instructions.  But it	*/
+	    /* seems to interfere enough with other cache traffic that the net	*/
+	    /* result is worse than prefetchnta.				*/
+#         if 0 
+	    /* Using prefetches for write seems to have a slight negative	*/
+	    /* impact on performance, at least for a PIII/500.			*/
+#	    define PREFETCH_FOR_WRITE(x) \
+	      __asm__ __volatile__ ("	prefetcht0	%0": : "m"(*(char *)(x)))
+#	  endif
+#	endif
+#	ifdef USE_3DNOW_PREFETCH
+#	  define PREFETCH(x) \
+	    __asm__ __volatile__ ("	prefetch	%0": : "m"(*(char *)(x)))
+#	  define PREFETCH_FOR_WRITE(x) \
+	    __asm__ __volatile__ ("	prefetchw	%0": : "m"(*(char *)(x)))
+#	endif
+#   endif
+#   ifdef CYGWIN32
+#       define OS_TYPE "CYGWIN32"
+#       define DATASTART ((ptr_t)GC_DATASTART)  /* From gc.h */
+#       define DATAEND	 ((ptr_t)GC_DATAEND)
+#	undef STACK_GRAN
+#       define STACK_GRAN 0x10000
+#       define HEURISTIC1
+#   endif
+#   ifdef OS2
+#	define OS_TYPE "OS2"
+ 	    	/* STACKBOTTOM and DATASTART are handled specially in 	*/
+		/* os_dep.c. OS2 actually has the right			*/
+		/* system call!						*/
+#	define DATAEND	/* not needed */
+#	define USE_GENERIC_PUSH_REGS
+#   endif
+#   ifdef MSWIN32
+#	define OS_TYPE "MSWIN32"
+		/* STACKBOTTOM and DATASTART are handled specially in 	*/
+		/* os_dep.c.						*/
+#       ifndef __WATCOMC__
+#	  define MPROTECT_VDB
+#	endif
+#       define DATAEND  /* not needed */
+#   endif
+#   ifdef MSWINCE
+#	define OS_TYPE "MSWINCE"
+#       define DATAEND  /* not needed */
+#   endif
+#   ifdef DJGPP
+#       define OS_TYPE "DJGPP"
+#       include "stubinfo.h"
+        extern int etext[];
+        extern int _stklen;
+        extern int __djgpp_stack_limit;
+#       define DATASTART ((ptr_t)((((word) (etext)) + 0x1ff) & ~0x1ff))
+/* #       define STACKBOTTOM ((ptr_t)((word) _stubinfo + _stubinfo->size \
+                                                     + _stklen)) */
+#       define STACKBOTTOM ((ptr_t)((word) __djgpp_stack_limit + _stklen))
+		/* This may not be right.  */
+#   endif
+#   ifdef OPENBSD
+#	define OS_TYPE "OPENBSD"
+#   endif
+#   ifdef FREEBSD
+#	define OS_TYPE "FREEBSD"
+#	ifndef GC_FREEBSD_THREADS
+#	    define MPROTECT_VDB
+#	endif
+#	ifdef __GLIBC__
+#	    define SIG_SUSPEND		(32+6)
+#	    define SIG_THR_RESTART	(32+5)
+	    extern int _end[];
+#	    define DATAEND (_end)
+#	else
+#	    define SIG_SUSPEND SIGUSR1
+#	    define SIG_THR_RESTART SIGUSR2
+#	endif
+#	define FREEBSD_STACKBOTTOM
+#	ifdef __ELF__
+#	    define DYNAMIC_LOADING
+#	endif
+	extern char etext[];
+	extern char * GC_FreeBSDGetDataStart();
+#	define DATASTART GC_FreeBSDGetDataStart(0x1000, &etext)
+#   endif
+#   ifdef NETBSD
+#	define OS_TYPE "NETBSD"
+#	ifdef __ELF__
+#	    define DYNAMIC_LOADING
+#	endif
+#   endif
+#   ifdef THREE86BSD
+#	define OS_TYPE "THREE86BSD"
+#   endif
+#   ifdef BSDI
+#	define OS_TYPE "BSDI"
+#   endif
+#   if defined(OPENBSD) || defined(NETBSD) \
+        || defined(THREE86BSD) || defined(BSDI)
+#	define HEURISTIC2
+	extern char etext[];
+#	define DATASTART ((ptr_t)(etext))
+#   endif
+#   ifdef NEXT
+#	define OS_TYPE "NEXT"
+#	define DATASTART ((ptr_t) get_etext())
+#	define STACKBOTTOM ((ptr_t)0xc0000000)
+#	define DATAEND	/* not needed */
+#   endif
+#   ifdef DOS4GW
+#     define OS_TYPE "DOS4GW"
+      extern long __nullarea;
+      extern char _end;
+      extern char *_STACKTOP;
+      /* Depending on calling conventions Watcom C either precedes
+         or does not precedes with undescore names of C-variables.
+         Make sure startup code variables always have the same names.  */
+      #pragma aux __nullarea "*";
+      #pragma aux _end "*";
+#     define STACKBOTTOM ((ptr_t) _STACKTOP)
+                         /* confused? me too. */
+#     define DATASTART ((ptr_t) &__nullarea)
+#     define DATAEND ((ptr_t) &_end)
+#   endif
+#   ifdef HURD
+#     define OS_TYPE "HURD"
+#     define STACK_GROWS_DOWN
+#     define HEURISTIC2
+      extern int  __data_start[];
+#     define DATASTART ( (ptr_t) (__data_start))
+      extern int   _end[];
+#     define DATAEND ( (ptr_t) (_end))
+/* #     define MPROTECT_VDB  Not quite working yet? */
+#     define DYNAMIC_LOADING
+#   endif
+#   ifdef DARWIN
+#     define OS_TYPE "DARWIN"
+#     define DARWIN_DONT_PARSE_STACK
+#     define DYNAMIC_LOADING
+      /* XXX: see get_end(3), get_etext() and get_end() should not be used.
+	 These aren't used when dyld support is enabled (it is by default) */
+#     define DATASTART ((ptr_t) get_etext())
+#     define DATAEND	((ptr_t) get_end())
+#     ifdef HAVE_PTHREAD_GET_STACKADDR_NP
+#       define STACKBOTTOM (ptr_t)pthread_get_stackaddr_np(pthread_self())
+#     else
+#       define STACKBOTTOM ((ptr_t) 0xc0000000)
+#     endif
+#     define USE_MMAP
+#     define USE_MMAP_ANON
+#     define USE_ASM_PUSH_REGS
+#     ifdef GC_DARWIN_THREADS
+#       define MPROTECT_VDB
+#     endif
+#     include <unistd.h>
+#     define GETPAGESIZE() getpagesize()
+      /* There seems to be some issues with trylock hanging on darwin. This
+	 should be looked into some more */
+#     define NO_PTHREAD_TRYLOCK
+#   endif /* DARWIN */
+# endif
+
+# ifdef NS32K
+#   define MACH_TYPE "NS32K"
+#   define ALIGNMENT 4
+    extern char **environ;
+#   define DATASTART ((ptr_t)(&environ))
+			      /* hideous kludge: environ is the first   */
+			      /* word in crt0.o, and delimits the start */
+			      /* of the data segment, no matter which   */
+			      /* ld options were passed through.        */
+#   define STACKBOTTOM ((ptr_t) 0xfffff000) /* for Encore */
+# endif
+
+# ifdef MIPS
+#   define MACH_TYPE "MIPS"
+#   ifdef LINUX
+      /* This was developed for a linuxce style platform.  Probably	*/
+      /* needs to be tweaked for workstation class machines.		*/
+#     define OS_TYPE "LINUX"
+#     define DYNAMIC_LOADING
+      extern int _end[];
+#     define DATAEND (_end)
+      extern int __data_start[];
+#     define DATASTART ((ptr_t)(__data_start))
+#     ifdef _MIPS_SZPTR
+#	define CPP_WORDSZ _MIPS_SZPTR
+#	define ALIGNMENT (_MIPS_SZPTR/8)
+#     else
+#	define ALIGNMENT 4
+#     endif
+#     ifndef HBLKSIZE
+#       define HBLKSIZE 4096
+#     endif
+#     define USE_GENERIC_PUSH_REGS
+#     if __GLIBC__ == 2 && __GLIBC_MINOR__ >= 2 || __GLIBC__ > 2
+#        define LINUX_STACKBOTTOM
+#     else
+#        define STACKBOTTOM 0x80000000
+#     endif
+#   endif /* Linux */
+#   ifdef EWS4800
+#      define HEURISTIC2
+#      if defined(_MIPS_SZPTR) && (_MIPS_SZPTR == 64)
+         extern int _fdata[], _end[];
+#        define DATASTART ((ptr_t)_fdata)
+#        define DATAEND ((ptr_t)_end)
+#        define CPP_WORDSZ _MIPS_SZPTR
+#        define ALIGNMENT (_MIPS_SZPTR/8)
+#      else
+         extern int etext[], edata[], end[];
+         extern int _DYNAMIC_LINKING[], _gp[];
+#        define DATASTART ((ptr_t)((((word)etext + 0x3ffff) & ~0x3ffff) \
+               + ((word)etext & 0xffff)))
+#        define DATAEND (edata)
+#        define DATASTART2 (_DYNAMIC_LINKING \
+               ? (ptr_t)(((word)_gp + 0x8000 + 0x3ffff) & ~0x3ffff) \
+               : (ptr_t)edata)
+#        define DATAEND2 (end)
+#        define ALIGNMENT 4
+#      endif
+#      define OS_TYPE "EWS4800"
+#      define USE_GENERIC_PUSH_REGS 1
+#   endif
+#   ifdef ULTRIX
+#	define HEURISTIC2
+#       define DATASTART (ptr_t)0x10000000
+			      /* Could probably be slightly higher since */
+			      /* startup code allocates lots of stuff.   */
+#	define OS_TYPE "ULTRIX"
+#       define ALIGNMENT 4
+#   endif
+#   ifdef RISCOS
+#	define HEURISTIC2
+#       define DATASTART (ptr_t)0x10000000
+#	define OS_TYPE "RISCOS"
+#   	define ALIGNMENT 4  /* Required by hardware */
+#   endif
+#   ifdef IRIX5
+#	define HEURISTIC2
+        extern int _fdata[];
+#       define DATASTART ((ptr_t)(_fdata))
+#       ifdef USE_MMAP
+#         define HEAP_START (ptr_t)0x30000000
+#       else
+#	  define HEAP_START DATASTART
+#       endif
+			      /* Lowest plausible heap address.		*/
+			      /* In the MMAP case, we map there.	*/
+			      /* In either case it is used to identify	*/
+			      /* heap sections so they're not 		*/
+			      /* considered as roots.			*/
+#	define OS_TYPE "IRIX5"
+/*#       define MPROTECT_VDB DOB: this should work, but there is evidence */
+/* 	  	of recent breakage.					   */
+#       ifdef _MIPS_SZPTR
+#	  define CPP_WORDSZ _MIPS_SZPTR
+#	  define ALIGNMENT (_MIPS_SZPTR/8)
+#	  if CPP_WORDSZ != 64
+#	    define ALIGN_DOUBLE
+#	  endif
+#	else
+#         define ALIGNMENT 4
+#	  define ALIGN_DOUBLE
+#	endif
+#	define DYNAMIC_LOADING
+#   endif
+#   ifdef MSWINCE
+#       define OS_TYPE "MSWINCE"
+#       define ALIGNMENT 4
+#       define DATAEND /* not needed */
+#   endif
+#   if defined(NETBSD)
+#     define ALIGNMENT 4
+#     define OS_TYPE "NETBSD"
+#     define HEURISTIC2
+#     define USE_GENERIC_PUSH_REGS
+#     ifdef __ELF__
+        extern int etext[];
+#       define DATASTART GC_data_start
+#       define NEED_FIND_LIMIT
+#       define DYNAMIC_LOADING
+#     else
+#       define DATASTART ((ptr_t) 0x10000000)
+#       define STACKBOTTOM ((ptr_t) 0x7ffff000)
+#     endif /* _ELF_ */
+#  endif
+# endif
+
+# ifdef RS6000
+#   define MACH_TYPE "RS6000"
+#   ifdef ALIGNMENT
+#     undef ALIGNMENT
+#   endif
+#   ifdef IA64
+#     undef IA64 /* DOB: some AIX installs stupidly define IA64 in /usr/include/sys/systemcfg.h */
+#   endif
+#   ifdef __64BIT__
+#     define ALIGNMENT 8
+#     define CPP_WORDSZ 64
+#     define STACKBOTTOM ((ptr_t)0x1000000000000000)
+#   else
+#     define ALIGNMENT 4
+#     define CPP_WORDSZ 32
+#     define STACKBOTTOM ((ptr_t)((ulong)&errno))
+#   endif
+#   define USE_MMAP
+#   define USE_MMAP_ANON
+ /* From AIX linker man page:
+ _text Specifies the first location of the program.
+ _etext Specifies the first location after the program.
+ _data Specifies the first location of the data.
+ _edata Specifies the first location after the initialized data
+ _end or end Specifies the first location after all data.
+ */
+    extern int _data[], _end[];
+#   define DATASTART ((ptr_t)((ulong)_data))
+#   define DATAEND ((ptr_t)((ulong)_end))
+    extern int errno;
+#   define USE_GENERIC_PUSH_REGS
+#   define DYNAMIC_LOADING
+	/* For really old versions of AIX, this may have to be removed. */
+# endif
+
+# ifdef HP_PA
+#   define MACH_TYPE "HP_PA"
+#   ifdef __LP64__
+#     define CPP_WORDSZ 64
+#     define ALIGNMENT 8
+#   else
+#     define CPP_WORDSZ 32
+#     define ALIGNMENT 4
+#     define ALIGN_DOUBLE
+#   endif
+#   if !defined(GC_HPUX_THREADS) && !defined(GC_LINUX_THREADS)
+#     ifndef LINUX /* For now. */
+#       define MPROTECT_VDB
+#     endif
+#   else
+#     define GENERIC_COMPARE_AND_SWAP
+	/* No compare-and-swap instruction.  Use pthread mutexes 	*/
+	/* when we absolutely have to.					*/
+#     ifdef PARALLEL_MARK
+#	define USE_MARK_BYTES
+		/* Minimize compare-and-swap usage.		*/
+#     endif
+#   endif
+#   define STACK_GROWS_UP
+#   ifdef HPUX
+#     define OS_TYPE "HPUX"
+      extern int __data_start[];
+#     define DATASTART ((ptr_t)(__data_start))
+#     if 0
+	/* The following appears to work for 7xx systems running HP/UX	*/
+	/* 9.xx Furthermore, it might result in much faster		*/
+	/* collections than HEURISTIC2, which may involve scanning	*/
+	/* segments that directly precede the stack.  It is not the	*/
+	/* default, since it may not work on older machine/OS		*/
+	/* combinations. (Thanks to Raymond X.T. Nijssen for uncovering	*/
+	/* this.)							*/
+#       define STACKBOTTOM ((ptr_t) 0x7b033000)  /* from /etc/conf/h/param.h */
+#     else
+	/* Gustavo Rodriguez-Rivera suggested changing HEURISTIC2	*/
+	/* to this.  Note that the GC must be initialized before the	*/
+    	/* first putenv call.						*/
+	extern char ** environ;
+#       define STACKBOTTOM ((ptr_t)environ)
+#     endif
+#     define DYNAMIC_LOADING
+#     include <unistd.h>
+#     define GETPAGESIZE() sysconf(_SC_PAGE_SIZE)
+#     ifndef __GNUC__
+#       define PREFETCH(x)  { \
+                              register long addr = (long)(x); \
+                              (void) _asm ("LDW", 0, 0, addr, 0); \
+                            }
+#     endif
+#   endif /* HPUX */
+#   ifdef LINUX
+#     define OS_TYPE "LINUX"
+#     define LINUX_STACKBOTTOM
+#     define DYNAMIC_LOADING
+#     define SEARCH_FOR_DATA_START
+      extern int _end[];
+#     define DATAEND (&_end)
+#   endif /* LINUX */
+# endif /* HP_PA */
+
+# ifdef ALPHA
+#   define MACH_TYPE "ALPHA"
+#   define ALIGNMENT 8
+#   define CPP_WORDSZ 64
+#   ifndef LINUX
+#     define USE_GENERIC_PUSH_REGS
+      /* Gcc and probably the DEC/Compaq compiler spill pointers to preserved */
+      /* fp registers in some cases when the target is a 21264.  The assembly */
+      /* code doesn't handle that yet, and version dependencies make that a   */
+      /* bit tricky.  Do the easy thing for now.				    */
+#   endif
+#   ifdef NETBSD
+#	define OS_TYPE "NETBSD"
+#	define HEURISTIC2
+#	define DATASTART GC_data_start
+#	define ELFCLASS32 32
+#	define ELFCLASS64 64
+#	define ELF_CLASS ELFCLASS64
+#       define DYNAMIC_LOADING
+#   endif
+#   ifdef OPENBSD
+#	define OS_TYPE "OPENBSD"
+#	define HEURISTIC2
+#   	ifdef __ELF__	/* since OpenBSD/Alpha 2.9 */
+#	   define DATASTART GC_data_start
+#   	   define ELFCLASS32 32
+#   	   define ELFCLASS64 64
+#   	   define ELF_CLASS ELFCLASS64
+#       else		/* ECOFF, until OpenBSD/Alpha 2.7 */
+#   	   define DATASTART ((ptr_t) 0x140000000)
+#   	endif
+#   endif
+#   ifdef FREEBSD
+#	define OS_TYPE "FREEBSD"
+/* MPROTECT_VDB is not yet supported at all on FreeBSD/alpha. */
+#	define SIG_SUSPEND SIGUSR1
+#	define SIG_THR_RESTART SIGUSR2
+#	define FREEBSD_STACKBOTTOM
+#	ifdef __ELF__
+#	    define DYNAMIC_LOADING
+#	endif
+/* Handle unmapped hole alpha*-*-freebsd[45]* puts between etext and edata. */
+	extern char etext[];
+	extern char edata[];
+	extern char end[];
+#	define NEED_FIND_LIMIT
+#	define DATASTART ((ptr_t)(&etext))
+#	define DATAEND (GC_find_limit (DATASTART, TRUE))
+#	define DATASTART2 ((ptr_t)(&edata))
+#	define DATAEND2 ((ptr_t)(&end))
+#   endif
+#   ifdef OSF1
+#	define OS_TYPE "OSF1"
+#   	define DATASTART ((ptr_t) 0x140000000)
+	extern int _end[];
+#   	define DATAEND ((ptr_t) &_end)
+ 	extern char ** environ;
+	/* round up from the value of environ to the nearest page boundary */
+	/* Probably breaks if putenv is called before collector 	   */
+	/* initialization.						   */
+#	define STACKBOTTOM ((ptr_t)(((word)(environ) | (getpagesize()-1))+1))
+/* #   	define HEURISTIC2 */
+	/* Normally HEURISTIC2 is too conervative, since		*/
+	/* the text segment immediately follows the stack.		*/
+	/* Hence we give an upper pound.				*/
+	/* This is currently unused, since we disabled HEURISTIC2	*/
+    	extern int __start[];
+#   	define HEURISTIC2_LIMIT ((ptr_t)((word)(__start) & ~(getpagesize()-1)))
+#	ifndef GC_OSF1_THREADS
+	  /* Unresolved signal issues with threads.	*/
+#   	  define MPROTECT_VDB
+#       endif
+#   	define DYNAMIC_LOADING
+#   endif
+#   ifdef LINUX
+#       define OS_TYPE "LINUX"
+#       define LINUX_STACKBOTTOM
+#       ifdef __ELF__
+#	  define SEARCH_FOR_DATA_START
+#         define DYNAMIC_LOADING
+#       else
+#           define DATASTART ((ptr_t) 0x140000000)
+#       endif
+	extern int _end[];
+#	define DATAEND (_end)
+#	define MPROTECT_VDB
+		/* Has only been superficially tested.  May not	*/
+		/* work on all versions.			*/
+#   endif
+# endif
+
+# ifdef IA64
+#   define MACH_TYPE "IA64"
+#   define USE_GENERIC_PUSH_REGS
+	/* We need to get preserved registers in addition to register   */
+	/* windows.   That's easiest to do with setjmp.			*/
+#   ifdef PARALLEL_MARK
+#	define USE_MARK_BYTES
+	    /* Compare-and-exchange is too expensive to use for 	*/
+	    /* setting mark bits.					*/
+#   endif
+#   ifdef HPUX
+#	ifdef _ILP32
+#	  define CPP_WORDSZ 32
+#         define ALIGN_DOUBLE
+	    /* Requires 8 byte alignment for malloc */
+#   	  define ALIGNMENT 4
+#       else
+#	  ifndef _LP64
+		---> unknown ABI
+#         endif
+#	  define CPP_WORDSZ 64
+#   	  define ALIGN_DOUBLE
+	    /* Requires 16 byte alignment for malloc */
+#         define ALIGNMENT 8
+#       endif
+#       define OS_TYPE "HPUX"	
+        extern int __data_start[];
+#       define DATASTART ((ptr_t)(__data_start))
+        /* Gustavo Rodriguez-Rivera suggested changing HEURISTIC2	*/
+        /* to this.  Note that the GC must be initialized before the	*/
+    	/* first putenv call.						*/
+	extern char ** environ;
+#       define STACKBOTTOM ((ptr_t)environ)
+#       define HPUX_STACKBOTTOM
+#       define DYNAMIC_LOADING
+#       include <unistd.h>
+#       define GETPAGESIZE() sysconf(_SC_PAGE_SIZE)
+ 	/* The following was empirically determined, and is probably	*/
+	/* not very robust.						*/
+	/* Note that the backing store base seems to be at a nice 	*/
+	/* address minus one page.					*/
+#	define BACKING_STORE_DISPLACEMENT 0x1000000
+#	define BACKING_STORE_ALIGNMENT 0x1000
+	extern ptr_t GC_register_stackbottom;
+#	define BACKING_STORE_BASE GC_register_stackbottom
+	/* Known to be wrong for recent HP/UX versions!!!	*/
+#   endif
+#   ifdef LINUX
+#   	define CPP_WORDSZ 64
+#   	define ALIGN_DOUBLE
+	  /* Requires 16 byte alignment for malloc */
+#   	define ALIGNMENT 8
+#       define OS_TYPE "LINUX"
+	/* The following works on NUE and older kernels:	*/
+/* #       define STACKBOTTOM ((ptr_t) 0xa000000000000000l)	*/
+	/* This does not work on NUE:				*/
+#       define LINUX_STACKBOTTOM
+	/* We also need the base address of the register stack	*/
+	/* backing store.  This is computed in 			*/
+	/* GC_linux_register_stack_base based on the following	*/
+	/* constants:						*/
+#       define BACKING_STORE_ALIGNMENT 0x100000
+#       define BACKING_STORE_DISPLACEMENT 0x80000000
+	extern ptr_t GC_register_stackbottom;
+#	define BACKING_STORE_BASE GC_register_stackbottom
+#	define SEARCH_FOR_DATA_START
+#	ifdef __GNUC__
+#         define DYNAMIC_LOADING
+#	else
+	  /* In the Intel compiler environment, we seem to end up with  */
+	  /* statically linked executables and an undefined reference	*/
+	  /* to _DYNAMIC						*/
+#  	endif
+#	define MPROTECT_VDB
+		/* Requires Linux 2.3.47 or later.	*/
+	extern int _end[];
+#	define DATAEND (_end)
+#       ifdef __GNUC__
+#	  ifndef __INTEL_COMPILER
+#	    define PREFETCH(x) \
+	      __asm__ ("	lfetch	[%0]": : "r"(x))
+#	    define PREFETCH_FOR_WRITE(x) \
+	      __asm__ ("	lfetch.excl	[%0]": : "r"(x))
+#	    define CLEAR_DOUBLE(x) \
+	      __asm__ ("	stf.spill	[%0]=f0": : "r"((void *)(x)))
+#	  else
+#           include <ia64intrin.h>
+#	    define PREFETCH(x) \
+	      __lfetch(__lfhint_none, (x))
+#	    define PREFETCH_FOR_WRITE(x) \
+	      __lfetch(__lfhint_nta,  (x))
+#	    define CLEAR_DOUBLE(x) \
+	      __stf_spill((void *)(x), 0)
+#	  endif // __INTEL_COMPILER
+#       endif
+#   endif
+#   ifdef MSWIN32
+      /* FIXME: This is a very partial guess.  There is no port, yet.	*/
+#     define OS_TYPE "MSWIN32"
+		/* STACKBOTTOM and DATASTART are handled specially in 	*/
+		/* os_dep.c.						*/
+#     define DATAEND  /* not needed */
+#     if defined(_WIN64)
+#       define CPP_WORDSZ 64
+#     else
+#       define CPP_WORDSZ 32   /* Is this possible?	*/
+#     endif
+#     define ALIGNMENT 8
+#   endif
+# endif
+
+# ifdef M88K
+#   define MACH_TYPE "M88K"
+#   define ALIGNMENT 4
+#   define ALIGN_DOUBLE
+    extern int etext[];
+#   ifdef CX_UX
+#	define OS_TYPE "CX_UX"
+#       define DATASTART ((((word)etext + 0x3fffff) & ~0x3fffff) + 0x10000)
+#   endif
+#   ifdef  DGUX
+#	define OS_TYPE "DGUX"
+	extern ptr_t GC_SysVGetDataStart();
+#       define DATASTART GC_SysVGetDataStart(0x10000, etext)
+#   endif
+#   define STACKBOTTOM ((char*)0xf0000000) /* determined empirically */
+# endif
+
+# ifdef S370
+    /* If this still works, and if anyone cares, this should probably	*/
+    /* be moved to the S390 category.					*/
+#   define MACH_TYPE "S370"
+#   define ALIGNMENT 4	/* Required by hardware	*/
+#   define USE_GENERIC_PUSH_REGS
+#   ifdef UTS4
+#       define OS_TYPE "UTS4"
+	extern int etext[];
+	extern int _etext[];
+	extern int _end[];
+	extern ptr_t GC_SysVGetDataStart();
+#       define DATASTART GC_SysVGetDataStart(0x10000, _etext)
+#	define DATAEND (_end)
+#	define HEURISTIC2
+#   endif
+# endif
+
+# ifdef S390
+#   define MACH_TYPE "S390"
+#   define USE_GENERIC_PUSH_REGS
+#   ifndef __s390x__
+#     define ALIGNMENT 4
+#     define CPP_WORDSZ 32
+#   else
+#     define ALIGNMENT 8
+#     define CPP_WORDSZ 64
+#   endif
+#   ifndef HBLKSIZE
+#     define HBLKSIZE 4096
+#   endif
+#   ifdef LINUX
+#       define OS_TYPE "LINUX"
+#       define LINUX_STACKBOTTOM
+#       define DYNAMIC_LOADING
+	extern int __data_start[];
+#       define DATASTART ((ptr_t)(__data_start))
+    extern int _end[];
+#   define DATAEND (_end)
+#   define CACHE_LINE_SIZE 256
+#   define GETPAGESIZE() 4096
+#   endif
+# endif
+
+# if defined(PJ)
+#   define ALIGNMENT 4
+    extern int _etext[];
+#   define DATASTART ((ptr_t)(_etext))
+#   define HEURISTIC1
+# endif
+
+# ifdef ARM32
+#   define CPP_WORDSZ 32
+#   define MACH_TYPE "ARM32"
+#   define ALIGNMENT 4
+#   ifdef NETBSD
+#       define OS_TYPE "NETBSD"
+#       define HEURISTIC2
+#	ifdef __ELF__
+#          define DATASTART GC_data_start
+#	   define DYNAMIC_LOADING
+#	else
+           extern char etext[];
+#          define DATASTART ((ptr_t)(etext))
+#	endif
+#       define USE_GENERIC_PUSH_REGS
+#   endif
+#   ifdef LINUX
+#       define OS_TYPE "LINUX"
+#       define LINUX_STACKBOTTOM
+#       undef STACK_GRAN
+#       define STACK_GRAN 0x10000000
+#       define USE_GENERIC_PUSH_REGS
+#       ifdef __ELF__
+#            define DYNAMIC_LOADING
+#	     include <features.h>
+#	     if defined(__GLIBC__) && __GLIBC__ >= 2
+#		 define SEARCH_FOR_DATA_START
+#	     else
+     	         extern char **__environ;
+#                define DATASTART ((ptr_t)(&__environ))
+			      /* hideous kludge: __environ is the first */
+			      /* word in crt0.o, and delimits the start */
+			      /* of the data segment, no matter which   */
+			      /* ld options were passed through.        */
+			      /* We could use _etext instead, but that  */
+			      /* would include .rodata, which may       */
+			      /* contain large read-only data tables    */
+			      /* that we'd rather not scan.		*/
+#	     endif
+	     extern int _end[];
+#	     define DATAEND (_end)
+#	else
+	     extern int etext[];
+#            define DATASTART ((ptr_t)((((word) (etext)) + 0xfff) & ~0xfff))
+#       endif
+#   endif
+#   ifdef MSWINCE
+#     define OS_TYPE "MSWINCE"
+#     define DATAEND /* not needed */
+#   endif
+#   ifdef NOSYS
+      /* __data_start is usually defined in the target linker script.  */
+      extern int __data_start[];
+#     define DATASTART (ptr_t)(__data_start)
+#     define USE_GENERIC_PUSH_REGS
+      /* __stack_base__ is set in newlib/libc/sys/arm/crt0.S  */
+      extern void *__stack_base__;
+#     define STACKBOTTOM ((ptr_t) (__stack_base__))
+#   endif
+#endif
+
+# ifdef CRIS
+#   define MACH_TYPE "CRIS"
+#   define CPP_WORDSZ 32
+#   define ALIGNMENT 1
+#   define OS_TYPE "LINUX"
+#   define DYNAMIC_LOADING
+#   define LINUX_STACKBOTTOM
+#   define USE_GENERIC_PUSH_REGS
+#   define SEARCH_FOR_DATA_START
+      extern int _end[];
+#   define DATAEND (_end)
+# endif
+
+# ifdef SH
+#   define MACH_TYPE "SH"
+#   define ALIGNMENT 4
+#   ifdef MSWINCE
+#     define OS_TYPE "MSWINCE"
+#     define DATAEND /* not needed */
+#   endif
+#   ifdef LINUX
+#     define OS_TYPE "LINUX"
+#     define LINUX_STACKBOTTOM
+#     define USE_GENERIC_PUSH_REGS
+#     define DYNAMIC_LOADING
+#     define SEARCH_FOR_DATA_START
+      extern int _end[];
+#     define DATAEND (_end)
+#   endif
+#   ifdef NETBSD
+#      define OS_TYPE "NETBSD"
+#      define HEURISTIC2
+#      define DATASTART GC_data_start
+#       define USE_GENERIC_PUSH_REGS
+#      define DYNAMIC_LOADING
+#   endif
+# endif
+ 
+# ifdef SH4
+#   define MACH_TYPE "SH4"
+#   define OS_TYPE "MSWINCE"
+#   define ALIGNMENT 4
+#   define DATAEND /* not needed */
+# endif
+
+# ifdef M32R
+#   define CPP_WORDSZ 32
+#   define MACH_TYPE "M32R"
+#   define ALIGNMENT 4
+#   ifdef LINUX
+#     define OS_TYPE "LINUX"
+#     define LINUX_STACKBOTTOM
+#     undef STACK_GRAN
+#     define STACK_GRAN 0x10000000
+#     define USE_GENERIC_PUSH_REGS
+#     define DYNAMIC_LOADING
+#     define SEARCH_FOR_DATA_START
+      extern int _end[];
+#     define DATAEND (_end)
+#   endif
+# endif
+
+# ifdef X86_64
+#   define MACH_TYPE "X86_64"
+#   define ALIGNMENT 8
+#   define CPP_WORDSZ 64
+#   ifndef HBLKSIZE
+#     define HBLKSIZE 4096
+#   endif
+#   define CACHE_LINE_SIZE 64
+#   define USE_GENERIC_PUSH_REGS
+#   ifdef LINUX
+#	define OS_TYPE "LINUX"
+#       define LINUX_STACKBOTTOM
+#       if !defined(GC_LINUX_THREADS) || !defined(REDIRECT_MALLOC)
+#	    define MPROTECT_VDB
+#	else
+	    /* We seem to get random errors in incremental mode,	*/
+	    /* possibly because Linux threads is itself a malloc client */
+	    /* and can't deal with the signals.				*/
+#	endif
+#       ifdef __ELF__
+#            define DYNAMIC_LOADING
+#	     ifdef UNDEFINED	/* includes ro data */
+	       extern int _etext[];
+#              define DATASTART ((ptr_t)((((word) (_etext)) + 0xfff) & ~0xfff))
+#	     endif
+#	     include <features.h>
+#	     define SEARCH_FOR_DATA_START
+	     extern int _end[];
+#	     define DATAEND (_end)
+#	else
+	     extern int etext[];
+#            define DATASTART ((ptr_t)((((word) (etext)) + 0xfff) & ~0xfff))
+#       endif
+#       if defined(__GNUC__) && __GNUC >= 3
+#	    define PREFETCH(x) __builtin_prefetch((x), 0, 0)
+#	    define PREFETCH_FOR_WRITE(x) __builtin_prefetch((x), 1)
+#	endif
+#   endif
+#   ifdef DARWIN
+#     define OS_TYPE "DARWIN"
+#     define DARWIN_DONT_PARSE_STACK
+#     define DYNAMIC_LOADING
+      /* XXX: see get_end(3), get_etext() and get_end() should not be used.
+	 These aren't used when dyld support is enabled (it is by default) */
+#     define DATASTART ((ptr_t) get_etext())
+#     define DATAEND	((ptr_t) get_end())
+#     ifdef HAVE_PTHREAD_GET_STACKADDR_NP
+#       define STACKBOTTOM (ptr_t)pthread_get_stackaddr_np(pthread_self())
+#     else
+#       define STACKBOTTOM ((ptr_t) 0x7fff5fc00000)
+#     endif
+#     define USE_MMAP
+#     define USE_MMAP_ANON
+#     ifdef GC_DARWIN_THREADS
+#       define MPROTECT_VDB
+#     endif
+#     include <unistd.h>
+#     define GETPAGESIZE() getpagesize()
+      /* There seems to be some issues with trylock hanging on darwin. This
+	 should be looked into some more */
+#     define NO_PTHREAD_TRYLOCK
+#   endif
+#   ifdef FREEBSD
+#	define OS_TYPE "FREEBSD"
+#	ifndef GC_FREEBSD_THREADS
+#	    define MPROTECT_VDB
+#	endif
+#	ifdef __GLIBC__
+#	    define SIG_SUSPEND		(32+6)
+#	    define SIG_THR_RESTART	(32+5)
+	    extern int _end[];
+#	    define DATAEND (_end)
+#	else
+#	    define SIG_SUSPEND SIGUSR1
+#	    define SIG_THR_RESTART SIGUSR2
+#	endif
+#	define FREEBSD_STACKBOTTOM
+#	ifdef __ELF__
+#	    define DYNAMIC_LOADING
+#	endif
+	extern char etext[];
+	extern char * GC_FreeBSDGetDataStart();
+#	define DATASTART GC_FreeBSDGetDataStart(0x1000, &etext)
+#   endif
+#   ifdef NETBSD
+#	define OS_TYPE "NETBSD"
+#	ifdef __ELF__
+#	    define DYNAMIC_LOADING
+#	endif
+#	define HEURISTIC2
+	extern char etext[];
+#	define SEARCH_FOR_DATA_START
+#   endif
+#   ifdef SUNOS5
+#	define ELF_CLASS ELFCLASS64
+#	define OS_TYPE "SUNOS5"
+        extern int _etext[], _end[];
+  	extern ptr_t GC_SysVGetDataStart();
+#       define DATASTART GC_SysVGetDataStart(0x1000, _etext)
+#	define DATAEND (_end)
+/*	# define STACKBOTTOM ((ptr_t)(_start)) worked through 2.7,  	*/
+/*      but reportedly breaks under 2.8.  It appears that the stack	*/
+/* 	base is a property of the executable, so this should not break	*/
+/* 	old executables.						*/
+/*  	HEURISTIC2 probably works, but this appears to be preferable.	*/
+/* #       include <sys/vm.h> */
+/* #	define STACKBOTTOM USRSTACK */
+#	define HEURISTIC2
+#	define PROC_VDB
+#	define DYNAMIC_LOADING
+#	if !defined(USE_MMAP) && defined(REDIRECT_MALLOC)
+#	    define USE_MMAP
+	    /* Otherwise we now use calloc.  Mmap may result in the	*/
+	    /* heap interleaved with thread stacks, which can result in	*/
+	    /* excessive blacklisting.  Sbrk is unusable since it	*/
+	    /* doesn't interact correctly with the system malloc.	*/
+#	endif
+#       ifdef USE_MMAP
+#         define HEAP_START (ptr_t)0x40000000
+#       else
+#	  define HEAP_START DATAEND
+#       endif
+#   endif
+# endif
+
+#if defined(LINUX) && defined(USE_MMAP)
+    /* The kernel may do a somewhat better job merging mappings etc.	*/
+    /* with anonymous mappings.						*/
+#   define USE_MMAP_ANON
+#endif
+
+#if defined(LINUX) && defined(REDIRECT_MALLOC)
+    /* Rld appears to allocate some memory with its own allocator, and	*/
+    /* some through malloc, which might be redirected.  To make this	*/
+    /* work with collectable memory, we have to scan memory allocated	*/
+    /* by rld's internal malloc.					*/
+#   define USE_PROC_FOR_LIBRARIES
+#endif
+    
+# ifndef STACK_GROWS_UP
+#   define STACK_GROWS_DOWN
+# endif
+
+# ifndef CPP_WORDSZ
+#   define CPP_WORDSZ 32
+# endif
+
+# ifndef OS_TYPE
+#   define OS_TYPE ""
+# endif
+
+# ifndef DATAEND
+    extern int end[];
+#   define DATAEND (end)
+# endif
+
+# if defined(SVR4) && !defined(GETPAGESIZE)
+#    include <unistd.h>
+#    define GETPAGESIZE()  sysconf(_SC_PAGESIZE)
+# endif
+
+# ifndef GETPAGESIZE
+#   if defined(SUNOS5) || defined(IRIX5)
+#	include <unistd.h>
+#   endif
+#   define GETPAGESIZE() getpagesize()
+# endif
+
+# if defined(SUNOS5) || defined(DRSNX) || defined(UTS4)
+	    /* OS has SVR4 generic features.  Probably others also qualify.	*/
+#   define SVR4
+# endif
+
+# if defined(SUNOS5) || defined(DRSNX)
+	    /* OS has SUNOS5 style semi-undocumented interface to dynamic 	*/
+	    /* loader.								*/
+#   define SUNOS5DL
+	    /* OS has SUNOS5 style signal handlers.				*/
+#   define SUNOS5SIGS
+# endif
+
+# if defined(HPUX)
+#   define SUNOS5SIGS
+# endif
+
+# if defined(FREEBSD) && ((__FreeBSD__ >= 4) || (__FreeBSD_kernel__ >= 4))
+#   define SUNOS5SIGS
+# endif
+
+# if defined(SVR4) || defined(LINUX) || defined(IRIX5) || defined(HPUX) \
+	    || defined(OPENBSD) || defined(NETBSD) || defined(FREEBSD) \
+	    || defined(DGUX) || defined(BSD) || defined(SUNOS4) \
+	    || defined(_AIX) || defined(DARWIN) || defined(OSF1)
+#   define UNIX_LIKE   /* Basic Unix-like system calls work.	*/
+# endif
+
+# if CPP_WORDSZ != 32 && CPP_WORDSZ != 64
+	   -> bad word size
+# endif
+
+# ifdef PCR
+#   undef DYNAMIC_LOADING
+#   undef STACKBOTTOM
+#   undef HEURISTIC1
+#   undef HEURISTIC2
+#   undef PROC_VDB
+#   undef MPROTECT_VDB
+#   define PCR_VDB
+# endif
+
+# ifdef SRC_M3
+	/* Postponed for now. */
+#   undef PROC_VDB
+#   undef MPROTECT_VDB
+# endif
+
+# ifdef SMALL_CONFIG
+	/* Presumably not worth the space it takes. */
+#   undef PROC_VDB
+#   undef MPROTECT_VDB
+# endif
+
+# ifdef USE_MUNMAP
+#   undef MPROTECT_VDB  /* Can't deal with address space holes. */
+# endif
+
+# ifdef PARALLEL_MARK
+#   undef MPROTECT_VDB  /* For now.	*/
+# endif
+
+# if !defined(PCR_VDB) && !defined(PROC_VDB) && !defined(MPROTECT_VDB)
+#   define DEFAULT_VDB
+# endif
+
+# ifndef PREFETCH
+#   define PREFETCH(x)
+#   define NO_PREFETCH
+# endif
+
+# ifndef PREFETCH_FOR_WRITE
+#   define PREFETCH_FOR_WRITE(x)
+#   define NO_PREFETCH_FOR_WRITE
+# endif
+
+# ifndef CACHE_LINE_SIZE
+#   define CACHE_LINE_SIZE 32	/* Wild guess	*/
+# endif
+
+# if defined(LINUX) || defined(__GLIBC__)
+#   define REGISTER_LIBRARIES_EARLY
+    /* We sometimes use dl_iterate_phdr, which may acquire an internal	*/
+    /* lock.  This isn't safe after the world has stopped.  So we must	*/
+    /* call GC_register_dynamic_libraries before stopping the world.	*/
+    /* For performance reasons, this may be beneficial on other		*/
+    /* platforms as well, though it should be avoided in win32.		*/
+# endif /* LINUX */
+
+# if defined(SEARCH_FOR_DATA_START)
+    extern ptr_t GC_data_start;
+#   define DATASTART GC_data_start
+# endif
+
+# ifndef CLEAR_DOUBLE
+#   define CLEAR_DOUBLE(x) \
+		((word*)x)[0] = 0; \
+		((word*)x)[1] = 0;
+# endif /* CLEAR_DOUBLE */
+
+# if defined(GC_IRIX_THREADS) && !defined(IRIX5)
+	--> inconsistent configuration
+# endif
+# if defined(GC_LINUX_THREADS) && !defined(LINUX)
+	--> inconsistent configuration
+# endif
+# if defined(GC_SOLARIS_THREADS) && !defined(SUNOS5)
+	--> inconsistent configuration
+# endif
+# if defined(GC_HPUX_THREADS) && !defined(HPUX)
+	--> inconsistent configuration
+# endif
+# if defined(GC_AIX_THREADS) && !defined(_AIX)
+	--> inconsistent configuration
+# endif
+# if defined(GC_WIN32_THREADS) && !defined(MSWIN32) && !defined(CYGWIN32)
+	--> inconsistent configuration
+# endif
+
+# if defined(PCR) || defined(SRC_M3) || \
+		defined(GC_SOLARIS_THREADS) || defined(GC_WIN32_THREADS) || \
+		defined(GC_PTHREADS)
+#   define THREADS
+# endif
+
+# if defined(HP_PA) || defined(M88K) \
+             || defined(POWERPC) && !defined(DARWIN) \
+	     || defined(LINT) || defined(MSWINCE) || defined(ARM32) || defined(CRIS) \
+	     || (defined(I386) && defined(__LCC__))
+	/* Use setjmp based hack to mark from callee-save registers.    */
+	/* The define should move to the individual platform 		*/
+	/* descriptions.						*/
+#	define USE_GENERIC_PUSH_REGS
+# endif
+
+# if defined(MSWINCE)
+#   define NO_GETENV
+# endif
+
+# if defined(SPARC)
+#   define ASM_CLEAR_CODE	/* Stack clearing is crucial, and we 	*/
+				/* include assembly code to do it well.	*/
+# endif
+
+  /* Can we save call chain in objects for debugging?   	        */
+  /* SET NFRAMES (# of saved frames) and NARGS (#of args for each 	*/
+  /* frame) to reasonable values for the platform.			*/
+  /* Set SAVE_CALL_CHAIN if we can.  SAVE_CALL_COUNT can be specified 	*/
+  /* at build time, though we feel free to adjust it slightly.		*/
+  /* Define NEED_CALLINFO if we either save the call stack or 		*/
+  /* GC_ADD_CALLER is defined.						*/
+  /* GC_CAN_SAVE_CALL_STACKS is set in gc.h.				*/
+
+#if defined(SPARC)
+# define CAN_SAVE_CALL_ARGS
+#endif
+#if (defined(I386) || defined(X86_64)) && (defined(LINUX) || defined(__GLIBC__))
+	    /* SAVE_CALL_CHAIN is supported if the code is compiled to save	*/
+	    /* frame pointers by default, i.e. no -fomit-frame-pointer flag.	*/
+# define CAN_SAVE_CALL_ARGS
+#endif
+
+# if defined(SAVE_CALL_COUNT) && !defined(GC_ADD_CALLER) \
+	     && defined(GC_CAN_SAVE_CALL_STACKS)
+#   define SAVE_CALL_CHAIN 
+# endif
+# ifdef SAVE_CALL_CHAIN
+#   if defined(SAVE_CALL_NARGS) && defined(CAN_SAVE_CALL_ARGS)
+#     define NARGS SAVE_CALL_NARGS
+#   else
+#     define NARGS 0	/* Number of arguments to save for each call.	*/
+#   endif
+# endif
+# ifdef SAVE_CALL_CHAIN
+#   ifndef SAVE_CALL_COUNT
+#     define NFRAMES 6	/* Number of frames to save. Even for		*/
+				/* alignment reasons.				*/
+#   else
+#     define NFRAMES ((SAVE_CALL_COUNT + 1) & ~1)
+#   endif
+#   define NEED_CALLINFO
+# endif /* SAVE_CALL_CHAIN */
+# ifdef GC_ADD_CALLER
+#   define NFRAMES 1
+#   define NARGS 0
+#   define NEED_CALLINFO
+# endif
+
+# if defined(MAKE_BACK_GRAPH) && !defined(DBG_HDRS_ALL)
+#   define DBG_HDRS_ALL
+# endif
+
+# if defined(POINTER_MASK) && !defined(POINTER_SHIFT)
+#   define POINTER_SHIFT 0
+# endif
+
+# if defined(POINTER_SHIFT) && !defined(POINTER_MASK)
+#   define POINTER_MASK ((GC_word)(-1))
+# endif
+
+# if !defined(FIXUP_POINTER) && defined(POINTER_MASK)
+#   define FIXUP_POINTER(p) (p) = ((p) & (POINTER_MASK) << POINTER_SHIFT)
+# endif
+
+# if defined(FIXUP_POINTER)
+#   define NEED_FIXUP_POINTER 1
+# else
+#   define NEED_FIXUP_POINTER 0
+#   define FIXUP_POINTER(p)
+# endif
+
+#ifdef GC_PRIVATE_H
+	/* This relies on some type definitions from gc_priv.h, from	*/
+        /* where it's normally included.				*/
+	/*								*/
+	/* How to get heap memory from the OS:				*/
+	/* Note that sbrk()-like allocation is preferred, since it 	*/
+	/* usually makes it possible to merge consecutively allocated	*/
+	/* chunks.  It also avoids unintented recursion with		*/
+	/* -DREDIRECT_MALLOC.						*/
+	/* GET_MEM() returns a HLKSIZE aligned chunk.			*/
+	/* 0 is taken to mean failure. 					*/
+	/* In the case os USE_MMAP, the argument must also be a 	*/
+	/* physical page size.						*/
+	/* GET_MEM is currently not assumed to retrieve 0 filled space, */
+	/* though we should perhaps take advantage of the case in which */
+	/* does.							*/
+	struct hblk;	/* See gc_priv.h.	*/
+# ifdef PCR
+	    char * real_malloc();
+#   define GET_MEM(bytes) HBLKPTR(real_malloc((size_t)bytes + GC_page_size) \
+					  + GC_page_size-1)
+# else
+#   ifdef OS2
+	      void * os2_alloc(size_t bytes);
+#     define GET_MEM(bytes) HBLKPTR((ptr_t)os2_alloc((size_t)bytes \
+					    + GC_page_size) \
+					    + GC_page_size-1)
+#   else
+#     if defined(NEXT) || defined(DOS4GW) || \
+		 (defined(AMIGA) && !defined(GC_AMIGA_FASTALLOC)) || \
+		 (defined(SUNOS5) && !defined(USE_MMAP))
+#       define GET_MEM(bytes) HBLKPTR((size_t) \
+					      calloc(1, (size_t)bytes + GC_page_size) \
+					      + GC_page_size-1)
+#     else
+#	if defined(MSWIN32) || defined(CYGWIN32)
+	  extern ptr_t GC_win32_get_mem();
+#         define GET_MEM(bytes) (struct hblk *)GC_win32_get_mem(bytes)
+#	else
+#	  ifdef MACOS
+#	    if defined(USE_TEMPORARY_MEMORY)
+			extern Ptr GC_MacTemporaryNewPtr(size_t size,
+							 Boolean clearMemory);
+#               define GET_MEM(bytes) HBLKPTR( \
+			    GC_MacTemporaryNewPtr(bytes + GC_page_size, true) \
+			    + GC_page_size-1)
+#	    else
+#         	    define GET_MEM(bytes) HBLKPTR( \
+				NewPtrClear(bytes + GC_page_size) + GC_page_size-1)
+#	    endif
+#	  else
+#	    ifdef MSWINCE
+	      extern ptr_t GC_wince_get_mem();
+#	      define GET_MEM(bytes) (struct hblk *)GC_wince_get_mem(bytes)
+#	    else
+#	      if defined(AMIGA) && defined(GC_AMIGA_FASTALLOC)
+			extern void *GC_amiga_get_mem(size_t size);
+#		define GET_MEM(bytes) HBLKPTR((size_t) \
+			  GC_amiga_get_mem((size_t)bytes + GC_page_size) \
+			  + GC_page_size-1)
+#	      else
+		extern ptr_t GC_unix_get_mem();
+#               define GET_MEM(bytes) (struct hblk *)GC_unix_get_mem(bytes)
+#	      endif
+#	    endif
+#	  endif
+#	endif
+#     endif
+#   endif
+# endif
+
+#endif /* GC_PRIVATE_H */
+
+#if defined(_AIX) && !defined(__GNUC__) && !defined(__STDC__)
+  /* IBMs xlc compiler doesn't appear to follow the convention of	*/
+  /* defining  __STDC__ to be zero in extended mode.			*/
+#   define __STDC__ 0
+#endif
+
+# endif /* GCCONFIG_H */
diff --git a/boehm-gc/include/private/pthread_stop_world.h b/boehm-gc/include/private/pthread_stop_world.h
new file mode 100644
index 000000000..6f9197a1f
--- /dev/null
+++ b/boehm-gc/include/private/pthread_stop_world.h
@@ -0,0 +1,11 @@
+#ifndef GC_PTHREAD_STOP_WORLD_H
+#define GC_PTHREAD_STOP_WORLD_H
+
+struct thread_stop_info {
+    word last_stop_count;	/* GC_last_stop_count value when thread	*/
+    				/* last successfully handled a suspend	*/
+    				/* signal.				*/
+    ptr_t stack_ptr;  		/* Valid only when stopped.      	*/
+};
+    
+#endif
diff --git a/boehm-gc/include/private/pthread_support.h b/boehm-gc/include/private/pthread_support.h
new file mode 100644
index 000000000..2186c0790
--- /dev/null
+++ b/boehm-gc/include/private/pthread_support.h
@@ -0,0 +1,103 @@
+#ifndef GC_PTHREAD_SUPPORT_H
+#define GC_PTHREAD_SUPPORT_H
+
+# include "private/gc_priv.h"
+
+# if defined(GC_PTHREADS) && !defined(GC_SOLARIS_THREADS) \
+     && !defined(GC_WIN32_THREADS)
+     
+#if defined(GC_DARWIN_THREADS)
+# include "private/darwin_stop_world.h"
+#else
+# include "private/pthread_stop_world.h"
+#endif
+
+/* We use the allocation lock to protect thread-related data structures. */
+
+/* The set of all known threads.  We intercept thread creation and 	*/
+/* joins.								*/
+/* Protected by allocation/GC lock.					*/
+/* Some of this should be declared volatile, but that's inconsistent	*/
+/* with some library routine declarations.  		 		*/
+typedef struct GC_Thread_Rep {
+    struct GC_Thread_Rep * next;  /* More recently allocated threads	*/
+				  /* with a given pthread id come 	*/
+				  /* first.  (All but the first are	*/
+				  /* guaranteed to be dead, but we may  */
+				  /* not yet have registered the join.) */
+    pthread_t id;
+    /* Extra bookkeeping information the stopping code uses */
+    struct thread_stop_info stop_info;
+    
+    short flags;
+#	define FINISHED 1   	/* Thread has exited.	*/
+#	define DETACHED 2	/* Thread is intended to be detached.	*/
+#	define MAIN_THREAD 4	/* True for the original thread only.	*/
+#       define SUSPENDED 8      /* True if thread was suspended externally */
+    short thread_blocked;	/* Protected by GC lock.		*/
+    				/* Treated as a boolean value.  If set,	*/
+    				/* thread will acquire GC lock before	*/
+    				/* doing any pointer manipulations, and	*/
+    				/* has set its sp value.  Thus it does	*/
+    				/* not need to be sent a signal to stop	*/
+    				/* it.					*/
+    ptr_t stack_end;		/* Cold end of the stack.		*/
+#   ifdef IA64
+	ptr_t backing_store_end;
+	ptr_t backing_store_ptr;
+#   endif
+    void * status;		/* The value returned from the thread.  */
+    				/* Used only to avoid premature 	*/
+				/* reclamation of any data it might 	*/
+				/* reference.				*/
+#   ifdef THREAD_LOCAL_ALLOC
+#	if CPP_WORDSZ == 64 && defined(ALIGN_DOUBLE)
+#	    define GRANULARITY 16
+#	    define NFREELISTS 49
+#	else
+#	    define GRANULARITY 8
+#	    define NFREELISTS 65
+#	endif
+	/* The ith free list corresponds to size i*GRANULARITY */
+#	define INDEX_FROM_BYTES(n) ((ADD_SLOP(n) + GRANULARITY - 1)/GRANULARITY)
+#	define BYTES_FROM_INDEX(i) ((i) * GRANULARITY - EXTRA_BYTES)
+#	define SMALL_ENOUGH(bytes) (ADD_SLOP(bytes) <= \
+				    (NFREELISTS-1)*GRANULARITY)
+	ptr_t ptrfree_freelists[NFREELISTS];
+	ptr_t normal_freelists[NFREELISTS];
+#	ifdef GC_GCJ_SUPPORT
+	  ptr_t gcj_freelists[NFREELISTS];
+#	endif
+		/* Free lists contain either a pointer or a small count */
+		/* reflecting the number of granules allocated at that	*/
+		/* size.						*/
+		/* 0 ==> thread-local allocation in use, free list	*/
+		/*       empty.						*/
+		/* > 0, <= DIRECT_GRANULES ==> Using global allocation,	*/
+		/*       too few objects of this size have been		*/
+		/* 	 allocated by this thread.			*/
+		/* >= HBLKSIZE  => pointer to nonempty free list.	*/
+		/* > DIRECT_GRANULES, < HBLKSIZE ==> transition to	*/
+		/*    local alloc, equivalent to 0.			*/
+#	define DIRECT_GRANULES (HBLKSIZE/GRANULARITY)
+		/* Don't use local free lists for up to this much 	*/
+		/* allocation.						*/
+#   endif
+} * GC_thread;
+
+# define THREAD_TABLE_SZ 128	/* Must be power of 2	*/
+extern volatile GC_thread GC_threads[THREAD_TABLE_SZ];
+
+extern GC_bool GC_thr_initialized;
+
+GC_thread GC_lookup_thread(pthread_t id);
+
+void GC_stop_init();
+
+extern GC_bool GC_in_thread_creation;
+	/* We may currently be in thread creation or destruction.	*/
+	/* Only set to TRUE while allocation lock is held.		*/
+	/* When set, it is OK to run GC from unknown thread.		*/
+
+#endif /* GC_PTHREADS && !GC_SOLARIS_THREADS.... etc */
+#endif /* GC_PTHREAD_SUPPORT_H */
diff --git a/boehm-gc/include/private/specific.h b/boehm-gc/include/private/specific.h
new file mode 100644
index 000000000..d04e19f5a
--- /dev/null
+++ b/boehm-gc/include/private/specific.h
@@ -0,0 +1,95 @@
+/*
+ * This is a reimplementation of a subset of the pthread_getspecific/setspecific
+ * interface. This appears to outperform the standard linuxthreads one
+ * by a significant margin.
+ * The major restriction is that each thread may only make a single
+ * pthread_setspecific call on a single key.  (The current data structure
+ * doesn't really require that.  The restriction should be easily removable.)
+ * We don't currently support the destruction functions, though that
+ * could be done.
+ * We also currently assume that only one pthread_setspecific call
+ * can be executed at a time, though that assumption would be easy to remove
+ * by adding a lock.
+ */
+
+#include <errno.h>
+
+/* Called during key creation or setspecific.		*/
+/* For the GC we already hold lock.			*/
+/* Currently allocated objects leak on thread exit.	*/
+/* That's hard to fix, but OK if we allocate garbage	*/
+/* collected memory.					*/
+#define MALLOC_CLEAR(n) GC_INTERNAL_MALLOC(n, NORMAL)
+#define PREFIXED(name) GC_##name
+
+#define TS_CACHE_SIZE 1024
+#define CACHE_HASH(n) (((((long)n) >> 8) ^ (long)n) & (TS_CACHE_SIZE - 1))
+#define TS_HASH_SIZE 1024
+#define HASH(n) (((((long)n) >> 8) ^ (long)n) & (TS_HASH_SIZE - 1))
+
+/* An entry describing a thread-specific value for a given thread.	*/
+/* All such accessible structures preserve the invariant that if either	*/
+/* thread is a valid pthread id or qtid is a valid "quick tread id"	*/
+/* for a thread, then value holds the corresponding thread specific	*/
+/* value.  This invariant must be preserved at ALL times, since		*/
+/* asynchronous reads are allowed.					*/
+typedef struct thread_specific_entry {
+	unsigned long qtid;	/* quick thread id, only for cache */
+	void * value;
+	struct thread_specific_entry *next;
+	pthread_t thread;
+} tse;
+
+
+/* We represent each thread-specific datum as two tables.  The first is	*/
+/* a cache, indexed by a "quick thread identifier".  The "quick" thread	*/
+/* identifier is an easy to compute value, which is guaranteed to	*/
+/* determine the thread, though a thread may correspond to more than	*/
+/* one value.  We typically use the address of a page in the stack.	*/
+/* The second is a hash table, indexed by pthread_self().  It is used	*/
+/* only as a backup.							*/
+
+/* Return the "quick thread id".  Default version.  Assumes page size,	*/
+/* or at least thread stack separation, is at least 4K.			*/
+/* Must be defined so that it never returns 0.  (Page 0 can't really	*/
+/* be part of any stack, since that would make 0 a valid stack pointer.)*/
+static __inline__ unsigned long quick_thread_id() {
+    int dummy;
+    return (unsigned long)(&dummy) >> 12;
+}
+
+#define INVALID_QTID ((unsigned long)0)
+#define INVALID_THREADID ((pthread_t)0)
+
+typedef struct thread_specific_data {
+    tse * volatile cache[TS_CACHE_SIZE];
+			/* A faster index to the hash table */
+    tse * hash[TS_HASH_SIZE];
+    pthread_mutex_t lock;
+} tsd;
+
+typedef tsd * PREFIXED(key_t);
+
+extern int PREFIXED(key_create) (tsd ** key_ptr, void (* destructor)(void *));
+
+extern int PREFIXED(setspecific) (tsd * key, void * value);
+
+extern void PREFIXED(remove_specific) (tsd * key);
+
+/* An internal version of getspecific that assumes a cache miss.	*/
+void * PREFIXED(slow_getspecific) (tsd * key, unsigned long qtid,
+				   tse * volatile * cache_entry);
+
+static __inline__ void * PREFIXED(getspecific) (tsd * key) {
+    long qtid = quick_thread_id();
+    unsigned hash_val = CACHE_HASH(qtid);
+    tse * volatile * entry_ptr = key -> cache + hash_val;
+    tse * entry = *entry_ptr;   /* Must be loaded only once.	*/
+    if (EXPECT(entry -> qtid == qtid, 1)) {
+      GC_ASSERT(entry -> thread == pthread_self());
+      return entry -> value;
+    }
+    return PREFIXED(slow_getspecific) (key, qtid, entry_ptr);
+}
+
+