implemented Setup.hs to build boehm cpp libs and install them;
[hs-boehmgc.git] / gc-7.2 / mark.c
diff --git a/gc-7.2/mark.c b/gc-7.2/mark.c
new file mode 100644 (file)
index 0000000..cc8dfde
--- /dev/null
@@ -0,0 +1,1852 @@
+/*
+ * Copyright 1988, 1989 Hans-J. Boehm, Alan J. Demers
+ * Copyright (c) 1991-1995 by Xerox Corporation.  All rights reserved.
+ * Copyright (c) 2000 by Hewlett-Packard Company.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
+ * OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program
+ * for any purpose,  provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is granted,
+ * provided the above notices are retained, and a notice that the code was
+ * modified is included with the above copyright notice.
+ *
+ */
+
+#include "private/gc_pmark.h"
+
+#include <stdio.h>
+
+#if defined(MSWIN32) && defined(__GNUC__)
+# include <excpt.h>
+#endif
+
+/* We put this here to minimize the risk of inlining. */
+/*VARARGS*/
+#if defined(__BORLANDC__) || defined(__WATCOMC__) || defined(__CC_ARM)
+  /*ARGSUSED*/
+  void GC_noop(void *p, ...) {}
+#else
+# ifdef __DMC__
+    void GC_noop(...) {}
+# else
+    void GC_noop() {}
+# endif
+#endif
+
+/* Single argument version, robust against whole program analysis. */
+volatile word GC_noop_sink;
+GC_API void GC_CALL GC_noop1(word x)
+{
+    GC_noop_sink = x;
+}
+
+/* mark_proc GC_mark_procs[MAX_MARK_PROCS] = {0} -- declared in gc_priv.h */
+
+GC_INNER unsigned GC_n_mark_procs = GC_RESERVED_MARK_PROCS;
+
+/* Initialize GC_obj_kinds properly and standard free lists properly.   */
+/* This must be done statically since they may be accessed before       */
+/* GC_init is called.                                                   */
+/* It's done here, since we need to deal with mark descriptors.         */
+GC_INNER struct obj_kind GC_obj_kinds[MAXOBJKINDS] = {
+/* PTRFREE */ { &GC_aobjfreelist[0], 0 /* filled in dynamically */,
+                0 | GC_DS_LENGTH, FALSE, FALSE },
+/* NORMAL  */ { &GC_objfreelist[0], 0,
+                0 | GC_DS_LENGTH,  /* Adjusted in GC_init for EXTRA_BYTES */
+                TRUE /* add length to descr */, TRUE },
+/* UNCOLLECTABLE */
+              { &GC_uobjfreelist[0], 0,
+                0 | GC_DS_LENGTH, TRUE /* add length to descr */, TRUE },
+# ifdef ATOMIC_UNCOLLECTABLE
+   /* AUNCOLLECTABLE */
+              { &GC_auobjfreelist[0], 0,
+                0 | GC_DS_LENGTH, FALSE /* add length to descr */, FALSE },
+# endif
+# ifdef STUBBORN_ALLOC
+/*STUBBORN*/ { (void **)&GC_sobjfreelist[0], 0,
+                0 | GC_DS_LENGTH, TRUE /* add length to descr */, TRUE },
+# endif
+};
+
+# ifdef ATOMIC_UNCOLLECTABLE
+#   ifdef STUBBORN_ALLOC
+#     define GC_N_KINDS_INITIAL_VALUE 5
+#   else
+#     define GC_N_KINDS_INITIAL_VALUE 4
+#   endif
+# else
+#   ifdef STUBBORN_ALLOC
+#     define GC_N_KINDS_INITIAL_VALUE 4
+#   else
+#     define GC_N_KINDS_INITIAL_VALUE 3
+#   endif
+# endif
+
+GC_INNER unsigned GC_n_kinds = GC_N_KINDS_INITIAL_VALUE;
+
+# ifndef INITIAL_MARK_STACK_SIZE
+#   define INITIAL_MARK_STACK_SIZE (1*HBLKSIZE)
+                /* INITIAL_MARK_STACK_SIZE * sizeof(mse) should be a    */
+                /* multiple of HBLKSIZE.                                */
+                /* The incremental collector actually likes a larger    */
+                /* size, since it want to push all marked dirty objs    */
+                /* before marking anything new.  Currently we let it    */
+                /* grow dynamically.                                    */
+# endif
+
+/*
+ * Limits of stack for GC_mark routine.
+ * All ranges between GC_mark_stack(incl.) and GC_mark_stack_top(incl.) still
+ * need to be marked from.
+ */
+
+STATIC word GC_n_rescuing_pages = 0;
+                                /* Number of dirty pages we marked from */
+                                /* excludes ptrfree pages, etc.         */
+
+GC_INNER mse * GC_mark_stack = NULL;
+GC_INNER mse * GC_mark_stack_limit = NULL;
+GC_INNER size_t GC_mark_stack_size = 0;
+
+#ifdef PARALLEL_MARK
+  GC_INNER mse * volatile GC_mark_stack_top = NULL;
+        /* Updated only with mark lock held, but read asynchronously.   */
+  STATIC volatile AO_t GC_first_nonempty = 0;
+        /* Lowest entry on mark stack   */
+        /* that may be nonempty.        */
+        /* Updated only by initiating   */
+        /* thread.                      */
+#else
+  GC_INNER mse * GC_mark_stack_top = NULL;
+#endif
+
+GC_INNER mark_state_t GC_mark_state = MS_NONE;
+
+GC_INNER GC_bool GC_mark_stack_too_small = FALSE;
+
+static struct hblk * scan_ptr;
+
+STATIC GC_bool GC_objects_are_marked = FALSE;
+                /* Are there collectable marked objects in the heap?    */
+
+/* Is a collection in progress?  Note that this can return true in the  */
+/* nonincremental case, if a collection has been abandoned and the      */
+/* mark state is now MS_INVALID.                                        */
+GC_INNER GC_bool GC_collection_in_progress(void)
+{
+    return(GC_mark_state != MS_NONE);
+}
+
+/* clear all mark bits in the header */
+GC_INNER void GC_clear_hdr_marks(hdr *hhdr)
+{
+    size_t last_bit = FINAL_MARK_BIT(hhdr -> hb_sz);
+    BZERO(hhdr -> hb_marks, sizeof(hhdr->hb_marks));
+    set_mark_bit_from_hdr(hhdr, last_bit);
+    hhdr -> hb_n_marks = 0;
+}
+
+/* Set all mark bits in the header.  Used for uncollectable blocks. */
+GC_INNER void GC_set_hdr_marks(hdr *hhdr)
+{
+    unsigned i;
+    size_t sz = hhdr -> hb_sz;
+    unsigned n_marks = (unsigned)FINAL_MARK_BIT(sz);
+
+#   ifdef USE_MARK_BYTES
+      for (i = 0; i <= n_marks; i += (unsigned)MARK_BIT_OFFSET(sz)) {
+        hhdr -> hb_marks[i] = 1;
+      }
+#   else
+      for (i = 0; i < divWORDSZ(n_marks + WORDSZ); ++i) {
+        hhdr -> hb_marks[i] = ONES;
+      }
+#   endif
+#   ifdef MARK_BIT_PER_OBJ
+      hhdr -> hb_n_marks = n_marks - 1;
+#   else
+      hhdr -> hb_n_marks = HBLK_OBJS(sz);
+#   endif
+}
+
+/*
+ * Clear all mark bits associated with block h.
+ */
+/*ARGSUSED*/
+static void clear_marks_for_block(struct hblk *h, word dummy)
+{
+    register hdr * hhdr = HDR(h);
+
+    if (IS_UNCOLLECTABLE(hhdr -> hb_obj_kind)) return;
+        /* Mark bit for these is cleared only once the object is        */
+        /* explicitly deallocated.  This either frees the block, or     */
+        /* the bit is cleared once the object is on the free list.      */
+    GC_clear_hdr_marks(hhdr);
+}
+
+/* Slow but general routines for setting/clearing/asking about mark bits */
+void GC_set_mark_bit(ptr_t p)
+{
+    struct hblk *h = HBLKPTR(p);
+    hdr * hhdr = HDR(h);
+    word bit_no = MARK_BIT_NO(p - (ptr_t)h, hhdr -> hb_sz);
+
+    if (!mark_bit_from_hdr(hhdr, bit_no)) {
+      set_mark_bit_from_hdr(hhdr, bit_no);
+      ++hhdr -> hb_n_marks;
+    }
+}
+
+void GC_clear_mark_bit(ptr_t p)
+{
+    struct hblk *h = HBLKPTR(p);
+    hdr * hhdr = HDR(h);
+    word bit_no = MARK_BIT_NO(p - (ptr_t)h, hhdr -> hb_sz);
+
+    if (mark_bit_from_hdr(hhdr, bit_no)) {
+      size_t n_marks;
+      clear_mark_bit_from_hdr(hhdr, bit_no);
+      n_marks = hhdr -> hb_n_marks - 1;
+#     ifdef PARALLEL_MARK
+        if (n_marks != 0 || !GC_parallel)
+          hhdr -> hb_n_marks = n_marks;
+        /* Don't decrement to zero.  The counts are approximate due to  */
+        /* concurrency issues, but we need to ensure that a count of    */
+        /* zero implies an empty block.                                 */
+#     else
+          hhdr -> hb_n_marks = n_marks;
+#     endif
+    }
+}
+
+GC_bool GC_is_marked(ptr_t p)
+{
+    struct hblk *h = HBLKPTR(p);
+    hdr * hhdr = HDR(h);
+    word bit_no = MARK_BIT_NO(p - (ptr_t)h, hhdr -> hb_sz);
+
+    return((GC_bool)mark_bit_from_hdr(hhdr, bit_no));
+}
+
+
+/*
+ * Clear mark bits in all allocated heap blocks.  This invalidates
+ * the marker invariant, and sets GC_mark_state to reflect this.
+ * (This implicitly starts marking to reestablish the invariant.)
+ */
+GC_INNER void GC_clear_marks(void)
+{
+    GC_apply_to_all_blocks(clear_marks_for_block, (word)0);
+    GC_objects_are_marked = FALSE;
+    GC_mark_state = MS_INVALID;
+    scan_ptr = 0;
+}
+
+#ifdef CHECKSUMS
+  void GC_check_dirty(void);
+#endif
+
+/* Initiate a garbage collection.  Initiates a full collection if the   */
+/* mark state is invalid.                                               */
+GC_INNER void GC_initiate_gc(void)
+{
+#   ifndef GC_DISABLE_INCREMENTAL
+        if (GC_dirty_maintained) GC_read_dirty();
+#   endif
+#   ifdef STUBBORN_ALLOC
+        GC_read_changed();
+#   endif
+#   ifdef CHECKSUMS
+        if (GC_dirty_maintained) GC_check_dirty();
+#   endif
+    GC_n_rescuing_pages = 0;
+    if (GC_mark_state == MS_NONE) {
+        GC_mark_state = MS_PUSH_RESCUERS;
+    } else if (GC_mark_state != MS_INVALID) {
+        ABORT("Unexpected state");
+    } /* else this is really a full collection, and mark        */
+      /* bits are invalid.                                      */
+    scan_ptr = 0;
+}
+
+#ifdef PARALLEL_MARK
+    STATIC void GC_do_parallel_mark(void); /* initiate parallel marking. */
+#endif /* PARALLEL_MARK */
+
+#ifdef GC_DISABLE_INCREMENTAL
+# define GC_push_next_marked_dirty(h) GC_push_next_marked(h)
+#else
+  STATIC struct hblk * GC_push_next_marked_dirty(struct hblk *h);
+                /* Invoke GC_push_marked on next dirty block above h.   */
+                /* Return a pointer just past the end of this block.    */
+#endif /* !GC_DISABLE_INCREMENTAL */
+STATIC struct hblk * GC_push_next_marked(struct hblk *h);
+                /* Ditto, but also mark from clean pages.       */
+STATIC struct hblk * GC_push_next_marked_uncollectable(struct hblk *h);
+                /* Ditto, but mark only from uncollectable pages.       */
+
+static void alloc_mark_stack(size_t);
+
+# if (defined(MSWIN32) || defined(MSWINCE)) && !defined(__GNUC__) \
+        || defined(MSWIN32) && defined(I386) /* for Win98 */ \
+        || defined(USE_PROC_FOR_LIBRARIES) && defined(THREADS)
+    /* Under rare conditions, we may end up marking from nonexistent memory. */
+    /* Hence we need to be prepared to recover by running GC_mark_some       */
+    /* with a suitable handler in place.                                     */
+    /* FIXME: Should we really need it for WinCE?  If yes then          */
+    /* WRAP_MARK_SOME should be also defined for CeGCC which requires   */
+    /* CPU/OS-specific code in mark_ex_handler() and GC_mark_some()     */
+    /* (for manual stack unwinding and exception handler installation). */
+#   define WRAP_MARK_SOME
+# endif
+
+/* Perform a small amount of marking.                   */
+/* We try to touch roughly a page of memory.            */
+/* Return TRUE if we just finished a mark phase.        */
+/* Cold_gc_frame is an address inside a GC frame that   */
+/* remains valid until all marking is complete.         */
+/* A zero value indicates that it's OK to miss some     */
+/* register values.                                     */
+/* We hold the allocation lock.  In the case of         */
+/* incremental collection, the world may not be stopped.*/
+#ifdef WRAP_MARK_SOME
+  /* For win32, this is called after we establish a structured  */
+  /* exception handler, in case Windows unmaps one of our root  */
+  /* segments.  See below.  In either case, we acquire the      */
+  /* allocator lock long before we get here.                    */
+  STATIC GC_bool GC_mark_some_inner(ptr_t cold_gc_frame)
+#else
+  GC_INNER GC_bool GC_mark_some(ptr_t cold_gc_frame)
+#endif
+{
+    switch(GC_mark_state) {
+        case MS_NONE:
+            return(FALSE);
+
+        case MS_PUSH_RESCUERS:
+            if (GC_mark_stack_top
+                >= GC_mark_stack_limit - INITIAL_MARK_STACK_SIZE/2) {
+                /* Go ahead and mark, even though that might cause us to */
+                /* see more marked dirty objects later on.  Avoid this   */
+                /* in the future.                                        */
+                GC_mark_stack_too_small = TRUE;
+                MARK_FROM_MARK_STACK();
+                return(FALSE);
+            } else {
+                scan_ptr = GC_push_next_marked_dirty(scan_ptr);
+                if (scan_ptr == 0) {
+                    if (GC_print_stats) {
+                        GC_log_printf("Marked from %lu dirty pages\n",
+                                      (unsigned long)GC_n_rescuing_pages);
+                    }
+                    GC_push_roots(FALSE, cold_gc_frame);
+                    GC_objects_are_marked = TRUE;
+                    if (GC_mark_state != MS_INVALID) {
+                        GC_mark_state = MS_ROOTS_PUSHED;
+                    }
+                }
+            }
+            return(FALSE);
+
+        case MS_PUSH_UNCOLLECTABLE:
+            if (GC_mark_stack_top
+                >= GC_mark_stack + GC_mark_stack_size/4) {
+#               ifdef PARALLEL_MARK
+                  /* Avoid this, since we don't parallelize the marker  */
+                  /* here.                                              */
+                  if (GC_parallel) GC_mark_stack_too_small = TRUE;
+#               endif
+                MARK_FROM_MARK_STACK();
+                return(FALSE);
+            } else {
+                scan_ptr = GC_push_next_marked_uncollectable(scan_ptr);
+                if (scan_ptr == 0) {
+                    GC_push_roots(TRUE, cold_gc_frame);
+                    GC_objects_are_marked = TRUE;
+                    if (GC_mark_state != MS_INVALID) {
+                        GC_mark_state = MS_ROOTS_PUSHED;
+                    }
+                }
+            }
+            return(FALSE);
+
+        case MS_ROOTS_PUSHED:
+#           ifdef PARALLEL_MARK
+              /* In the incremental GC case, this currently doesn't     */
+              /* quite do the right thing, since it runs to             */
+              /* completion.  On the other hand, starting a             */
+              /* parallel marker is expensive, so perhaps it is         */
+              /* the right thing?                                       */
+              /* Eventually, incremental marking should run             */
+              /* asynchronously in multiple threads, without grabbing   */
+              /* the allocation lock.                                   */
+                if (GC_parallel) {
+                  GC_do_parallel_mark();
+                  GC_ASSERT(GC_mark_stack_top < (mse *)GC_first_nonempty);
+                  GC_mark_stack_top = GC_mark_stack - 1;
+                  if (GC_mark_stack_too_small) {
+                    alloc_mark_stack(2*GC_mark_stack_size);
+                  }
+                  if (GC_mark_state == MS_ROOTS_PUSHED) {
+                    GC_mark_state = MS_NONE;
+                    return(TRUE);
+                  } else {
+                    return(FALSE);
+                  }
+                }
+#           endif
+            if (GC_mark_stack_top >= GC_mark_stack) {
+                MARK_FROM_MARK_STACK();
+                return(FALSE);
+            } else {
+                GC_mark_state = MS_NONE;
+                if (GC_mark_stack_too_small) {
+                    alloc_mark_stack(2*GC_mark_stack_size);
+                }
+                return(TRUE);
+            }
+
+        case MS_INVALID:
+        case MS_PARTIALLY_INVALID:
+            if (!GC_objects_are_marked) {
+                GC_mark_state = MS_PUSH_UNCOLLECTABLE;
+                return(FALSE);
+            }
+            if (GC_mark_stack_top >= GC_mark_stack) {
+                MARK_FROM_MARK_STACK();
+                return(FALSE);
+            }
+            if (scan_ptr == 0 && GC_mark_state == MS_INVALID) {
+                /* About to start a heap scan for marked objects. */
+                /* Mark stack is empty.  OK to reallocate.        */
+                if (GC_mark_stack_too_small) {
+                    alloc_mark_stack(2*GC_mark_stack_size);
+                }
+                GC_mark_state = MS_PARTIALLY_INVALID;
+            }
+            scan_ptr = GC_push_next_marked(scan_ptr);
+            if (scan_ptr == 0 && GC_mark_state == MS_PARTIALLY_INVALID) {
+                GC_push_roots(TRUE, cold_gc_frame);
+                GC_objects_are_marked = TRUE;
+                if (GC_mark_state != MS_INVALID) {
+                    GC_mark_state = MS_ROOTS_PUSHED;
+                }
+            }
+            return(FALSE);
+        default:
+            ABORT("GC_mark_some: bad state");
+            return(FALSE);
+    }
+}
+
+#ifdef WRAP_MARK_SOME
+
+# if (defined(MSWIN32) || defined(MSWINCE)) && defined(__GNUC__)
+
+    typedef struct {
+      EXCEPTION_REGISTRATION ex_reg;
+      void *alt_path;
+    } ext_ex_regn;
+
+    static EXCEPTION_DISPOSITION mark_ex_handler(
+        struct _EXCEPTION_RECORD *ex_rec,
+        void *est_frame,
+        struct _CONTEXT *context,
+        void *disp_ctxt)
+    {
+        if (ex_rec->ExceptionCode == STATUS_ACCESS_VIOLATION) {
+          ext_ex_regn *xer = (ext_ex_regn *)est_frame;
+
+          /* Unwind from the inner function assuming the standard */
+          /* function prologue.                                   */
+          /* Assumes code has not been compiled with              */
+          /* -fomit-frame-pointer.                                */
+          context->Esp = context->Ebp;
+          context->Ebp = *((DWORD *)context->Esp);
+          context->Esp = context->Esp - 8;
+
+          /* Resume execution at the "real" handler within the    */
+          /* wrapper function.                                    */
+          context->Eip = (DWORD )(xer->alt_path);
+
+          return ExceptionContinueExecution;
+
+        } else {
+            return ExceptionContinueSearch;
+        }
+    }
+# endif /* __GNUC__  && MSWIN32 */
+
+#if defined(GC_WIN32_THREADS) && !defined(__GNUC__)
+  GC_bool GC_started_thread_while_stopped(void);
+  /* In win32_threads.c.  Did we invalidate mark phase with an  */
+  /* unexpected thread start?                                   */
+#endif
+
+  GC_INNER GC_bool GC_mark_some(ptr_t cold_gc_frame)
+  {
+      GC_bool ret_val;
+
+#   if defined(MSWIN32) || defined(MSWINCE)
+#    ifndef __GNUC__
+      /* Windows 98 appears to asynchronously create and remove  */
+      /* writable memory mappings, for reasons we haven't yet    */
+      /* understood.  Since we look for writable regions to      */
+      /* determine the root set, we may try to mark from an      */
+      /* address range that disappeared since we started the     */
+      /* collection.  Thus we have to recover from faults here.  */
+      /* This code does not appear to be necessary for Windows   */
+      /* 95/NT/2000+. Note that this code should never generate  */
+      /* an incremental GC write fault.                          */
+      /* This code seems to be necessary for WinCE (at least in  */
+      /* the case we'd decide to add MEM_PRIVATE sections to     */
+      /* data roots in GC_register_dynamic_libraries()).         */
+      /* It's conceivable that this is the same issue with       */
+      /* terminating threads that we see with Linux and          */
+      /* USE_PROC_FOR_LIBRARIES.                                 */
+
+      __try {
+          ret_val = GC_mark_some_inner(cold_gc_frame);
+      } __except (GetExceptionCode() == EXCEPTION_ACCESS_VIOLATION ?
+                EXCEPTION_EXECUTE_HANDLER : EXCEPTION_CONTINUE_SEARCH) {
+          goto handle_ex;
+      }
+#     ifdef GC_WIN32_THREADS
+        /* With DllMain-based thread tracking, a thread may have        */
+        /* started while we were marking.  This is logically equivalent */
+        /* to the exception case; our results are invalid and we have   */
+        /* to start over.  This cannot be prevented since we can't      */
+        /* block in DllMain.                                            */
+        if (GC_started_thread_while_stopped()) goto handle_ex;
+#     endif
+     rm_handler:
+      return ret_val;
+
+#    else /* __GNUC__ */
+
+      /* Manually install an exception handler since GCC does    */
+      /* not yet support Structured Exception Handling (SEH) on  */
+      /* Win32.                                                  */
+
+      ext_ex_regn er;
+
+      er.alt_path = &&handle_ex;
+      er.ex_reg.handler = mark_ex_handler;
+      __asm__ __volatile__ ("movl %%fs:0, %0" : "=r" (er.ex_reg.prev));
+      __asm__ __volatile__ ("movl %0, %%fs:0" : : "r" (&er));
+      ret_val = GC_mark_some_inner(cold_gc_frame);
+      /* Prevent GCC from considering the following code unreachable */
+      /* and thus eliminating it.                                    */
+        if (er.alt_path == 0)
+          goto handle_ex;
+    rm_handler:
+      /* Uninstall the exception handler */
+      __asm__ __volatile__ ("mov %0, %%fs:0" : : "r" (er.ex_reg.prev));
+      return ret_val;
+
+#    endif /* __GNUC__ */
+#   else /* !MSWIN32 */
+      /* Here we are handling the case in which /proc is used for root  */
+      /* finding, and we have threads.  We may find a stack for a       */
+      /* thread that is in the process of exiting, and disappears       */
+      /* while we are marking it.  This seems extremely difficult to    */
+      /* avoid otherwise.                                               */
+      if (GC_incremental) {
+        WARN("Incremental GC incompatible with /proc roots\n", 0);
+        /* I'm not sure if this could still work ...    */
+      }
+      GC_setup_temporary_fault_handler();
+      if(SETJMP(GC_jmp_buf) != 0) goto handle_ex;
+      ret_val = GC_mark_some_inner(cold_gc_frame);
+    rm_handler:
+      GC_reset_fault_handler();
+      return ret_val;
+
+#   endif /* !MSWIN32 */
+
+handle_ex:
+    /* Exception handler starts here for all cases. */
+      if (GC_print_stats) {
+        GC_log_printf(
+          "Caught ACCESS_VIOLATION in marker; memory mapping disappeared\n");
+      }
+
+      /* We have bad roots on the stack.  Discard mark stack.   */
+      /* Rescan from marked objects.  Redetermine roots.        */
+      GC_invalidate_mark_state();
+      scan_ptr = 0;
+
+      ret_val = FALSE;
+      goto rm_handler;  /* Back to platform-specific code. */
+  }
+#endif /* WRAP_MARK_SOME */
+
+GC_INNER GC_bool GC_mark_stack_empty(void)
+{
+    return(GC_mark_stack_top < GC_mark_stack);
+}
+
+GC_INNER void GC_invalidate_mark_state(void)
+{
+    GC_mark_state = MS_INVALID;
+    GC_mark_stack_top = GC_mark_stack-1;
+}
+
+GC_INNER mse * GC_signal_mark_stack_overflow(mse *msp)
+{
+    GC_mark_state = MS_INVALID;
+    GC_mark_stack_too_small = TRUE;
+    if (GC_print_stats) {
+        GC_log_printf("Mark stack overflow; current size = %lu entries\n",
+                      (unsigned long)GC_mark_stack_size);
+    }
+    return(msp - GC_MARK_STACK_DISCARDS);
+}
+
+/*
+ * Mark objects pointed to by the regions described by
+ * mark stack entries between mark_stack and mark_stack_top,
+ * inclusive.  Assumes the upper limit of a mark stack entry
+ * is never 0.  A mark stack entry never has size 0.
+ * We try to traverse on the order of a hblk of memory before we return.
+ * Caller is responsible for calling this until the mark stack is empty.
+ * Note that this is the most performance critical routine in the
+ * collector.  Hence it contains all sorts of ugly hacks to speed
+ * things up.  In particular, we avoid procedure calls on the common
+ * path, we take advantage of peculiarities of the mark descriptor
+ * encoding, we optionally maintain a cache for the block address to
+ * header mapping, we prefetch when an object is "grayed", etc.
+ */
+GC_INNER mse * GC_mark_from(mse *mark_stack_top, mse *mark_stack,
+                            mse *mark_stack_limit)
+{
+  signed_word credit = HBLKSIZE;  /* Remaining credit for marking work  */
+  ptr_t current_p;      /* Pointer to current candidate ptr.    */
+  word current; /* Candidate pointer.                   */
+  ptr_t limit;  /* (Incl) limit of current candidate    */
+                                /* range                                */
+  word descr;
+  ptr_t greatest_ha = GC_greatest_plausible_heap_addr;
+  ptr_t least_ha = GC_least_plausible_heap_addr;
+  DECLARE_HDR_CACHE;
+
+# define SPLIT_RANGE_WORDS 128  /* Must be power of 2.          */
+
+  GC_objects_are_marked = TRUE;
+  INIT_HDR_CACHE;
+# ifdef OS2 /* Use untweaked version to circumvent compiler problem */
+    while (mark_stack_top >= mark_stack && credit >= 0)
+# else
+    while ((((ptr_t)mark_stack_top - (ptr_t)mark_stack) | credit) >= 0)
+# endif
+  {
+    current_p = mark_stack_top -> mse_start;
+    descr = mark_stack_top -> mse_descr;
+  retry:
+    /* current_p and descr describe the current object.         */
+    /* *mark_stack_top is vacant.                               */
+    /* The following is 0 only for small objects described by a simple  */
+    /* length descriptor.  For many applications this is the common     */
+    /* case, so we try to detect it quickly.                            */
+    if (descr & ((~(WORDS_TO_BYTES(SPLIT_RANGE_WORDS) - 1)) | GC_DS_TAGS)) {
+      word tag = descr & GC_DS_TAGS;
+
+      switch(tag) {
+        case GC_DS_LENGTH:
+          /* Large length.                                              */
+          /* Process part of the range to avoid pushing too much on the */
+          /* stack.                                                     */
+          GC_ASSERT(descr < (word)GC_greatest_plausible_heap_addr
+                            - (word)GC_least_plausible_heap_addr);
+#         ifdef ENABLE_TRACE
+            if (GC_trace_addr >= current_p
+                && GC_trace_addr < current_p + descr) {
+              GC_log_printf("GC:%u Large section; start %p len %lu\n",
+                        (unsigned)GC_gc_no, current_p, (unsigned long)descr);
+            }
+#         endif /* ENABLE_TRACE */
+#         ifdef PARALLEL_MARK
+#           define SHARE_BYTES 2048
+            if (descr > SHARE_BYTES && GC_parallel
+                && mark_stack_top < mark_stack_limit - 1) {
+              int new_size = (descr/2) & ~(sizeof(word)-1);
+              mark_stack_top -> mse_start = current_p;
+              mark_stack_top -> mse_descr = new_size + sizeof(word);
+                                        /* makes sure we handle         */
+                                        /* misaligned pointers.         */
+              mark_stack_top++;
+#             ifdef ENABLE_TRACE
+                if (GC_trace_addr >= current_p
+                    && GC_trace_addr < current_p + descr) {
+                  GC_log_printf("GC:%u Splitting (parallel) %p at %p\n",
+                        (unsigned)GC_gc_no, current_p, current_p + new_size);
+                }
+#             endif /* ENABLE_TRACE */
+              current_p += new_size;
+              descr -= new_size;
+              goto retry;
+            }
+#         endif /* PARALLEL_MARK */
+          mark_stack_top -> mse_start =
+                limit = current_p + WORDS_TO_BYTES(SPLIT_RANGE_WORDS-1);
+          mark_stack_top -> mse_descr =
+                        descr - WORDS_TO_BYTES(SPLIT_RANGE_WORDS-1);
+#         ifdef ENABLE_TRACE
+            if (GC_trace_addr >= current_p
+                && GC_trace_addr < current_p + descr) {
+              GC_log_printf("GC:%u Splitting %p at %p\n",
+                            (unsigned)GC_gc_no, current_p, limit);
+            }
+#         endif /* ENABLE_TRACE */
+          /* Make sure that pointers overlapping the two ranges are     */
+          /* considered.                                                */
+          limit += sizeof(word) - ALIGNMENT;
+          break;
+        case GC_DS_BITMAP:
+          mark_stack_top--;
+#         ifdef ENABLE_TRACE
+            if (GC_trace_addr >= current_p
+                && GC_trace_addr < current_p + WORDS_TO_BYTES(WORDSZ-2)) {
+              GC_log_printf("GC:%u Tracing from %p bitmap descr %lu\n",
+                        (unsigned)GC_gc_no, current_p, (unsigned long)descr);
+            }
+#         endif /* ENABLE_TRACE */
+          descr &= ~GC_DS_TAGS;
+          credit -= WORDS_TO_BYTES(WORDSZ/2); /* guess */
+          while (descr != 0) {
+            if ((signed_word)descr < 0) {
+              current = *(word *)current_p;
+              FIXUP_POINTER(current);
+              if ((ptr_t)current >= least_ha && (ptr_t)current < greatest_ha) {
+                PREFETCH((ptr_t)current);
+#               ifdef ENABLE_TRACE
+                  if (GC_trace_addr == current_p) {
+                    GC_log_printf("GC:%u Considering(3) %p -> %p\n",
+                               (unsigned)GC_gc_no, current_p, (ptr_t)current);
+                  }
+#               endif /* ENABLE_TRACE */
+                PUSH_CONTENTS((ptr_t)current, mark_stack_top,
+                              mark_stack_limit, current_p, exit1);
+              }
+            }
+            descr <<= 1;
+            current_p += sizeof(word);
+          }
+          continue;
+        case GC_DS_PROC:
+          mark_stack_top--;
+#         ifdef ENABLE_TRACE
+            if (GC_trace_addr >= current_p
+                && GC_base(current_p) != 0
+                && GC_base(current_p) == GC_base(GC_trace_addr)) {
+              GC_log_printf("GC:%u Tracing from %p proc descr %lu\n",
+                        (unsigned)GC_gc_no, current_p, (unsigned long)descr);
+            }
+#         endif /* ENABLE_TRACE */
+          credit -= GC_PROC_BYTES;
+          mark_stack_top =
+              (*PROC(descr))
+                    ((word *)current_p, mark_stack_top,
+                    mark_stack_limit, ENV(descr));
+          continue;
+        case GC_DS_PER_OBJECT:
+          if ((signed_word)descr >= 0) {
+            /* Descriptor is in the object.     */
+            descr = *(word *)(current_p + descr - GC_DS_PER_OBJECT);
+          } else {
+            /* Descriptor is in type descriptor pointed to by first     */
+            /* word in object.                                          */
+            ptr_t type_descr = *(ptr_t *)current_p;
+            /* type_descr is either a valid pointer to the descriptor   */
+            /* structure, or this object was on a free list.  If it     */
+            /* it was anything but the last object on the free list,    */
+            /* we will misinterpret the next object on the free list as */
+            /* the type descriptor, and get a 0 GC descriptor, which    */
+            /* is ideal.  Unfortunately, we need to check for the last  */
+            /* object case explicitly.                                  */
+            if (0 == type_descr) {
+                /* Rarely executed.     */
+                mark_stack_top--;
+                continue;
+            }
+            descr = *(word *)(type_descr
+                              - (descr + (GC_INDIR_PER_OBJ_BIAS
+                                          - GC_DS_PER_OBJECT)));
+          }
+          if (0 == descr) {
+              /* Can happen either because we generated a 0 descriptor  */
+              /* or we saw a pointer to a free object.          */
+              mark_stack_top--;
+              continue;
+          }
+          goto retry;
+        default:
+          /* Can't happen. */
+          limit = 0; /* initialized to prevent warning. */
+      }
+    } else /* Small object with length descriptor */ {
+      mark_stack_top--;
+      limit = current_p + (word)descr;
+    }
+#   ifdef ENABLE_TRACE
+        if (GC_trace_addr >= current_p
+            && GC_trace_addr < limit) {
+          GC_log_printf("GC:%u Tracing from %p len %lu\n",
+                        (int)GC_gc_no, current_p, (unsigned long)descr);
+        }
+#   endif /* ENABLE_TRACE */
+    /* The simple case in which we're scanning a range. */
+    GC_ASSERT(!((word)current_p & (ALIGNMENT-1)));
+    credit -= limit - current_p;
+    limit -= sizeof(word);
+    {
+#     define PREF_DIST 4
+
+#     ifndef SMALL_CONFIG
+        word deferred;
+
+        /* Try to prefetch the next pointer to be examined asap.        */
+        /* Empirically, this also seems to help slightly without        */
+        /* prefetches, at least on linux/X86.  Presumably this loop     */
+        /* ends up with less register pressure, and gcc thus ends up    */
+        /* generating slightly better code.  Overall gcc code quality   */
+        /* for this loop is still not great.                            */
+        for(;;) {
+          PREFETCH(limit - PREF_DIST*CACHE_LINE_SIZE);
+          GC_ASSERT(limit >= current_p);
+          deferred = *(word *)limit;
+          FIXUP_POINTER(deferred);
+          limit -= ALIGNMENT;
+          if ((ptr_t)deferred >= least_ha && (ptr_t)deferred <  greatest_ha) {
+            PREFETCH((ptr_t)deferred);
+            break;
+          }
+          if (current_p > limit) goto next_object;
+          /* Unroll once, so we don't do too many of the prefetches     */
+          /* based on limit.                                            */
+          deferred = *(word *)limit;
+          FIXUP_POINTER(deferred);
+          limit -= ALIGNMENT;
+          if ((ptr_t)deferred >= least_ha && (ptr_t)deferred <  greatest_ha) {
+            PREFETCH((ptr_t)deferred);
+            break;
+          }
+          if (current_p > limit) goto next_object;
+        }
+#     endif
+
+      while (current_p <= limit) {
+        /* Empirically, unrolling this loop doesn't help a lot. */
+        /* Since PUSH_CONTENTS expands to a lot of code,        */
+        /* we don't.                                            */
+        current = *(word *)current_p;
+        FIXUP_POINTER(current);
+        PREFETCH(current_p + PREF_DIST*CACHE_LINE_SIZE);
+        if ((ptr_t)current >= least_ha && (ptr_t)current <  greatest_ha) {
+          /* Prefetch the contents of the object we just pushed.  It's  */
+          /* likely we will need them soon.                             */
+          PREFETCH((ptr_t)current);
+#         ifdef ENABLE_TRACE
+            if (GC_trace_addr == current_p) {
+              GC_log_printf("GC:%u Considering(1) %p -> %p\n",
+                            (unsigned)GC_gc_no, current_p, (ptr_t)current);
+            }
+#         endif /* ENABLE_TRACE */
+          PUSH_CONTENTS((ptr_t)current, mark_stack_top,
+                           mark_stack_limit, current_p, exit2);
+        }
+        current_p += ALIGNMENT;
+      }
+
+#     ifndef SMALL_CONFIG
+        /* We still need to mark the entry we previously prefetched.    */
+        /* We already know that it passes the preliminary pointer       */
+        /* validity test.                                               */
+#       ifdef ENABLE_TRACE
+            if (GC_trace_addr == current_p) {
+              GC_log_printf("GC:%u Considering(2) %p -> %p\n",
+                            (unsigned)GC_gc_no, current_p, (ptr_t)deferred);
+            }
+#       endif /* ENABLE_TRACE */
+        PUSH_CONTENTS((ptr_t)deferred, mark_stack_top,
+                         mark_stack_limit, current_p, exit4);
+        next_object:;
+#     endif
+    }
+  }
+  return mark_stack_top;
+}
+
+#ifdef PARALLEL_MARK
+
+STATIC GC_bool GC_help_wanted = FALSE;  /* Protected by mark lock       */
+STATIC unsigned GC_helper_count = 0;    /* Number of running helpers.   */
+                                        /* Protected by mark lock       */
+STATIC unsigned GC_active_count = 0;    /* 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_INNER word GC_mark_no = 0;
+
+#define LOCAL_MARK_STACK_SIZE HBLKSIZE
+        /* Under normal circumstances, this is big enough to guarantee  */
+        /* We don't overflow half of it in a single call to             */
+        /* GC_mark_from.                                                */
+
+
+/* Steal mark stack entries starting at mse low into mark stack local   */
+/* until we either steal mse high, or we have max entries.              */
+/* Return a pointer to the top of the local mark stack.                 */
+/* *next is replaced by a pointer to the next unscanned mark stack      */
+/* entry.                                                               */
+STATIC mse * GC_steal_mark_stack(mse * low, mse * high, mse * local,
+                                 unsigned max, mse **next)
+{
+    mse *p;
+    mse *top = local - 1;
+    unsigned i = 0;
+
+    GC_ASSERT(high >= low-1 && (word)(high - low + 1) <= GC_mark_stack_size);
+    for (p = low; p <= high && i <= max; ++p) {
+        word descr = AO_load((volatile AO_t *) &(p -> mse_descr));
+        if (descr != 0) {
+            /* Must be ordered after read of descr: */
+            AO_store_release_write((volatile AO_t *) &(p -> mse_descr), 0);
+            /* More than one thread may get this entry, but that's only */
+            /* a minor performance problem.                             */
+            ++top;
+            top -> mse_descr = descr;
+            top -> mse_start = p -> mse_start;
+            GC_ASSERT((top -> mse_descr & GC_DS_TAGS) != GC_DS_LENGTH ||
+                      top -> mse_descr < (word)GC_greatest_plausible_heap_addr
+                                         - (word)GC_least_plausible_heap_addr);
+            /* If this is a big object, count it as                     */
+            /* size/256 + 1 objects.                                    */
+            ++i;
+            if ((descr & GC_DS_TAGS) == GC_DS_LENGTH) i += (int)(descr >> 8);
+        }
+    }
+    *next = p;
+    return top;
+}
+
+/* Copy back a local mark stack.        */
+/* low and high are inclusive bounds.   */
+STATIC void GC_return_mark_stack(mse * low, mse * high)
+{
+    mse * my_top;
+    mse * my_start;
+    size_t stack_size;
+
+    if (high < low) return;
+    stack_size = high - low + 1;
+    GC_acquire_mark_lock();
+    my_top = GC_mark_stack_top; /* Concurrent modification impossible. */
+    my_start = my_top + 1;
+    if (my_start - GC_mark_stack + stack_size > GC_mark_stack_size) {
+      if (GC_print_stats) {
+          GC_log_printf("No room to copy back mark stack\n");
+      }
+      GC_mark_state = MS_INVALID;
+      GC_mark_stack_too_small = TRUE;
+      /* We drop the local mark stack.  We'll fix things later. */
+    } else {
+      BCOPY(low, my_start, stack_size * sizeof(mse));
+      GC_ASSERT((mse *)AO_load((volatile AO_t *)(&GC_mark_stack_top))
+                == my_top);
+      AO_store_release_write((volatile AO_t *)(&GC_mark_stack_top),
+                             (AO_t)(my_top + stack_size));
+                /* Ensures visibility of previously written stack contents. */
+    }
+    GC_release_mark_lock();
+    GC_notify_all_marker();
+}
+
+/* Mark from the local mark stack.              */
+/* On return, the local mark stack is empty.    */
+/* But this may be achieved by copying the      */
+/* local mark stack back into the global one.   */
+STATIC void GC_do_local_mark(mse *local_mark_stack, mse *local_top)
+{
+    unsigned n;
+#   define N_LOCAL_ITERS 1
+
+#   ifdef GC_ASSERTIONS
+      /* Make sure we don't hold mark lock. */
+        GC_acquire_mark_lock();
+        GC_release_mark_lock();
+#   endif
+    for (;;) {
+        for (n = 0; n < N_LOCAL_ITERS; ++n) {
+            local_top = GC_mark_from(local_top, local_mark_stack,
+                                     local_mark_stack + LOCAL_MARK_STACK_SIZE);
+            if (local_top < local_mark_stack) return;
+            if ((word)(local_top - local_mark_stack)
+                        >= LOCAL_MARK_STACK_SIZE / 2) {
+                GC_return_mark_stack(local_mark_stack, local_top);
+                return;
+            }
+        }
+        if ((mse *)AO_load((volatile AO_t *)(&GC_mark_stack_top))
+            < (mse *)AO_load(&GC_first_nonempty)
+            && GC_active_count < GC_helper_count
+            && local_top > local_mark_stack + 1) {
+            /* Try to share the load, since the main stack is empty,    */
+            /* and helper threads are waiting for a refill.             */
+            /* The entries near the bottom of the stack are likely      */
+            /* to require more work.  Thus we return those, eventhough  */
+            /* it's harder.                                             */
+            mse * new_bottom = local_mark_stack
+                                + (local_top - local_mark_stack)/2;
+            GC_ASSERT(new_bottom > local_mark_stack
+                      && new_bottom < local_top);
+            GC_return_mark_stack(local_mark_stack, new_bottom - 1);
+            memmove(local_mark_stack, new_bottom,
+                    (local_top - new_bottom + 1) * sizeof(mse));
+            local_top -= (new_bottom - local_mark_stack);
+        }
+    }
+}
+
+#define ENTRIES_TO_GET 5
+
+GC_INNER long GC_markers = 2;   /* Normally changed by thread-library-  */
+                                /* -specific code.                      */
+
+/* Mark using the local mark stack until the global mark stack is empty */
+/* and there are no active workers. Update GC_first_nonempty to reflect */
+/* progress.                                                            */
+/* Caller does not hold mark lock.                                      */
+/* Caller has already incremented GC_helper_count.  We decrement it,    */
+/* and maintain GC_active_count.                                        */
+STATIC void GC_mark_local(mse *local_mark_stack, int id)
+{
+    mse * my_first_nonempty;
+
+    GC_acquire_mark_lock();
+    GC_active_count++;
+    my_first_nonempty = (mse *)AO_load(&GC_first_nonempty);
+    GC_ASSERT((mse *)AO_load(&GC_first_nonempty) >= GC_mark_stack &&
+              (mse *)AO_load(&GC_first_nonempty) <=
+              (mse *)AO_load((volatile AO_t *)(&GC_mark_stack_top)) + 1);
+    if (GC_print_stats == VERBOSE)
+        GC_log_printf("Starting mark helper %lu\n", (unsigned long)id);
+    GC_release_mark_lock();
+    for (;;) {
+        size_t n_on_stack;
+        unsigned n_to_get;
+        mse * my_top;
+        mse * local_top;
+        mse * global_first_nonempty = (mse *)AO_load(&GC_first_nonempty);
+
+        GC_ASSERT(my_first_nonempty >= GC_mark_stack &&
+                  my_first_nonempty <=
+                  (mse *)AO_load((volatile AO_t *)(&GC_mark_stack_top))  + 1);
+        GC_ASSERT(global_first_nonempty >= GC_mark_stack &&
+                  global_first_nonempty <=
+                  (mse *)AO_load((volatile AO_t *)(&GC_mark_stack_top))  + 1);
+        if (my_first_nonempty < global_first_nonempty) {
+            my_first_nonempty = global_first_nonempty;
+        } else if (global_first_nonempty < my_first_nonempty) {
+            AO_compare_and_swap(&GC_first_nonempty,
+                                (AO_t) global_first_nonempty,
+                                (AO_t) my_first_nonempty);
+            /* If this fails, we just go ahead, without updating        */
+            /* GC_first_nonempty.                                       */
+        }
+        /* Perhaps we should also update GC_first_nonempty, if it */
+        /* is less.  But that would require using atomic updates. */
+        my_top = (mse *)AO_load_acquire((volatile AO_t *)(&GC_mark_stack_top));
+        n_on_stack = my_top - my_first_nonempty + 1;
+        if (0 == n_on_stack) {
+            GC_acquire_mark_lock();
+            my_top = GC_mark_stack_top;
+                /* Asynchronous modification impossible here,   */
+                /* since we hold mark lock.                     */
+            n_on_stack = my_top - my_first_nonempty + 1;
+            if (0 == n_on_stack) {
+                GC_active_count--;
+                GC_ASSERT(GC_active_count <= GC_helper_count);
+                /* Other markers may redeposit objects  */
+                /* on the stack.                                */
+                if (0 == GC_active_count) GC_notify_all_marker();
+                while (GC_active_count > 0
+                       && (mse *)AO_load(&GC_first_nonempty)
+                          > GC_mark_stack_top) {
+                    /* We will be notified if either GC_active_count    */
+                    /* reaches zero, or if more objects are pushed on   */
+                    /* the global mark stack.                           */
+                    GC_wait_marker();
+                }
+                if (GC_active_count == 0 &&
+                    (mse *)AO_load(&GC_first_nonempty) > GC_mark_stack_top) {
+                    GC_bool need_to_notify = FALSE;
+                    /* The above conditions can't be falsified while we */
+                    /* hold the mark lock, since neither                */
+                    /* GC_active_count nor GC_mark_stack_top can        */
+                    /* change.  GC_first_nonempty can only be           */
+                    /* incremented asynchronously.  Thus we know that   */
+                    /* both conditions actually held simultaneously.    */
+                    GC_helper_count--;
+                    if (0 == GC_helper_count) need_to_notify = TRUE;
+                    if (GC_print_stats == VERBOSE)
+                      GC_log_printf("Finished mark helper %lu\n",
+                                    (unsigned long)id);
+                    GC_release_mark_lock();
+                    if (need_to_notify) GC_notify_all_marker();
+                    return;
+                }
+                /* else there's something on the stack again, or        */
+                /* another helper may push something.                   */
+                GC_active_count++;
+                GC_ASSERT(GC_active_count > 0);
+                GC_release_mark_lock();
+                continue;
+            } else {
+                GC_release_mark_lock();
+            }
+        }
+        n_to_get = ENTRIES_TO_GET;
+        if (n_on_stack < 2 * ENTRIES_TO_GET) n_to_get = 1;
+        local_top = GC_steal_mark_stack(my_first_nonempty, my_top,
+                                        local_mark_stack, n_to_get,
+                                        &my_first_nonempty);
+        GC_ASSERT(my_first_nonempty >= GC_mark_stack &&
+                  my_first_nonempty <=
+                    (mse *)AO_load((volatile AO_t *)(&GC_mark_stack_top)) + 1);
+        GC_do_local_mark(local_mark_stack, local_top);
+    }
+}
+
+/* Perform Parallel mark.                       */
+/* We hold the GC lock, not the mark lock.      */
+/* Currently runs until the mark stack is       */
+/* empty.                                       */
+STATIC void GC_do_parallel_mark(void)
+{
+    mse local_mark_stack[LOCAL_MARK_STACK_SIZE];
+
+    GC_acquire_mark_lock();
+    GC_ASSERT(I_HOLD_LOCK());
+    /* This could be a GC_ASSERT, but it seems safer to keep it on      */
+    /* all the time, especially since it's cheap.                       */
+    if (GC_help_wanted || GC_active_count != 0 || GC_helper_count != 0)
+        ABORT("Tried to start parallel mark in bad state");
+    if (GC_print_stats == VERBOSE)
+        GC_log_printf("Starting marking for mark phase number %lu\n",
+                      (unsigned long)GC_mark_no);
+    GC_first_nonempty = (AO_t)GC_mark_stack;
+    GC_active_count = 0;
+    GC_helper_count = 1;
+    GC_help_wanted = TRUE;
+    GC_release_mark_lock();
+    GC_notify_all_marker();
+        /* Wake up potential helpers.   */
+    GC_mark_local(local_mark_stack, 0);
+    GC_acquire_mark_lock();
+    GC_help_wanted = FALSE;
+    /* Done; clean up.  */
+    while (GC_helper_count > 0) GC_wait_marker();
+    /* GC_helper_count cannot be incremented while GC_help_wanted == FALSE */
+    if (GC_print_stats == VERBOSE)
+        GC_log_printf("Finished marking for mark phase number %lu\n",
+                      (unsigned long)GC_mark_no);
+    GC_mark_no++;
+    GC_release_mark_lock();
+    GC_notify_all_marker();
+}
+
+
+/* Try to help out the marker, if it's running.         */
+/* We do not hold the GC lock, but the requestor does.  */
+GC_INNER void GC_help_marker(word my_mark_no)
+{
+    mse local_mark_stack[LOCAL_MARK_STACK_SIZE];
+    unsigned my_id;
+
+    if (!GC_parallel) return;
+    GC_acquire_mark_lock();
+    while (GC_mark_no < my_mark_no
+           || (!GC_help_wanted && GC_mark_no == my_mark_no)) {
+      GC_wait_marker();
+    }
+    my_id = GC_helper_count;
+    if (GC_mark_no != my_mark_no || my_id >= (unsigned)GC_markers) {
+      /* Second test is useful only if original threads can also        */
+      /* act as helpers.  Under Linux they can't.                       */
+      GC_release_mark_lock();
+      return;
+    }
+    GC_helper_count = my_id + 1;
+    GC_release_mark_lock();
+    GC_mark_local(local_mark_stack, my_id);
+    /* GC_mark_local decrements GC_helper_count. */
+}
+
+#endif /* PARALLEL_MARK */
+
+/* Allocate or reallocate space for mark stack of size n entries.  */
+/* May silently fail.                                              */
+static void alloc_mark_stack(size_t n)
+{
+    mse * new_stack = (mse *)GC_scratch_alloc(n * sizeof(struct GC_ms_entry));
+#   ifdef GWW_VDB
+      /* Don't recycle a stack segment obtained with the wrong flags.   */
+      /* Win32 GetWriteWatch requires the right kind of memory.         */
+      static GC_bool GC_incremental_at_stack_alloc = FALSE;
+      GC_bool recycle_old = (!GC_incremental || GC_incremental_at_stack_alloc);
+
+      GC_incremental_at_stack_alloc = GC_incremental;
+#   else
+#     define recycle_old TRUE
+#   endif
+
+    GC_mark_stack_too_small = FALSE;
+    if (GC_mark_stack_size != 0) {
+        if (new_stack != 0) {
+          if (recycle_old) {
+            /* Recycle old space */
+              size_t page_offset = (word)GC_mark_stack & (GC_page_size - 1);
+              size_t size = GC_mark_stack_size * sizeof(struct GC_ms_entry);
+              size_t displ = 0;
+
+              if (0 != page_offset) displ = GC_page_size - page_offset;
+              size = (size - displ) & ~(GC_page_size - 1);
+              if (size > 0) {
+                GC_add_to_heap((struct hblk *)
+                                ((word)GC_mark_stack + displ), (word)size);
+              }
+          }
+          GC_mark_stack = new_stack;
+          GC_mark_stack_size = n;
+          GC_mark_stack_limit = new_stack + n;
+          if (GC_print_stats) {
+              GC_log_printf("Grew mark stack to %lu frames\n",
+                            (unsigned long) GC_mark_stack_size);
+          }
+        } else {
+          if (GC_print_stats) {
+              GC_log_printf("Failed to grow mark stack to %lu frames\n",
+                            (unsigned long) n);
+          }
+        }
+    } else {
+        if (new_stack == 0) {
+            GC_err_printf("No space for mark stack\n");
+            EXIT();
+        }
+        GC_mark_stack = new_stack;
+        GC_mark_stack_size = n;
+        GC_mark_stack_limit = new_stack + n;
+    }
+    GC_mark_stack_top = GC_mark_stack-1;
+}
+
+GC_INNER void GC_mark_init(void)
+{
+    alloc_mark_stack(INITIAL_MARK_STACK_SIZE);
+}
+
+/*
+ * Push all locations between b and t onto the mark stack.
+ * b is the first location to be checked. t is one past the last
+ * location to be checked.
+ * Should only be used if there is no possibility of mark stack
+ * overflow.
+ */
+void GC_push_all(ptr_t bottom, ptr_t top)
+{
+    register word length;
+
+    bottom = (ptr_t)(((word) bottom + ALIGNMENT-1) & ~(ALIGNMENT-1));
+    top = (ptr_t)(((word) top) & ~(ALIGNMENT-1));
+    if (bottom >= top) return;
+
+    GC_mark_stack_top++;
+    if (GC_mark_stack_top >= GC_mark_stack_limit) {
+        ABORT("Unexpected mark stack overflow");
+    }
+    length = top - bottom;
+#   if GC_DS_TAGS > ALIGNMENT - 1
+        length += GC_DS_TAGS;
+        length &= ~GC_DS_TAGS;
+#   endif
+    GC_mark_stack_top -> mse_start = bottom;
+    GC_mark_stack_top -> mse_descr = length;
+}
+
+#ifndef GC_DISABLE_INCREMENTAL
+
+  /* Analogous to the above, but push only those pages h with           */
+  /* dirty_fn(h) != 0.  We use GC_push_all to actually push the block.  */
+  /* Used both to selectively push dirty pages, or to push a block in   */
+  /* piecemeal fashion, to allow for more marking concurrency.          */
+  /* Will not overflow mark stack if GC_push_all pushes a small fixed   */
+  /* number of entries.  (This is invoked only if GC_push_all pushes    */
+  /* a single entry, or if it marks each object before pushing it, thus */
+  /* ensuring progress in the event of a stack overflow.)               */
+  STATIC void GC_push_selected(ptr_t bottom, ptr_t top,
+                               GC_bool (*dirty_fn)(struct hblk *))
+  {
+    struct hblk * h;
+
+    bottom = (ptr_t)(((word) bottom + ALIGNMENT-1) & ~(ALIGNMENT-1));
+    top = (ptr_t)(((word) top) & ~(ALIGNMENT-1));
+    if (bottom >= top) return;
+
+    h = HBLKPTR(bottom + HBLKSIZE);
+    if (top <= (ptr_t) h) {
+        if ((*dirty_fn)(h-1)) {
+            GC_push_all(bottom, top);
+        }
+        return;
+    }
+    if ((*dirty_fn)(h-1)) {
+        GC_push_all(bottom, (ptr_t)h);
+    }
+
+    while ((ptr_t)(h+1) <= top) {
+        if ((*dirty_fn)(h)) {
+            if ((word)(GC_mark_stack_top - GC_mark_stack)
+                > 3 * GC_mark_stack_size / 4) {
+                /* Danger of mark stack overflow */
+                GC_push_all((ptr_t)h, top);
+                return;
+            } else {
+                GC_push_all((ptr_t)h, (ptr_t)(h+1));
+            }
+        }
+        h++;
+    }
+
+    if ((ptr_t)h != top && (*dirty_fn)(h)) {
+       GC_push_all((ptr_t)h, top);
+    }
+    if (GC_mark_stack_top >= GC_mark_stack_limit) {
+        ABORT("Unexpected mark stack overflow");
+    }
+  }
+
+  void GC_push_conditional(ptr_t bottom, ptr_t top, GC_bool all)
+  {
+    if (!all) {
+      GC_push_selected(bottom, top, GC_page_was_dirty);
+    } else {
+#     ifdef PROC_VDB
+        if (GC_dirty_maintained) {
+          /* Pages that were never dirtied cannot contain pointers.     */
+          GC_push_selected(bottom, top, GC_page_was_ever_dirty);
+        } else
+#     endif
+      /* else */ {
+        GC_push_all(bottom, top);
+      }
+    }
+  }
+#endif /* !GC_DISABLE_INCREMENTAL */
+
+#if defined(MSWIN32) || defined(MSWINCE)
+  void __cdecl GC_push_one(word p)
+#else
+  void GC_push_one(word p)
+#endif
+{
+    GC_PUSH_ONE_STACK(p, MARKED_FROM_REGISTER);
+}
+
+/*ARGSUSED*/
+GC_API struct GC_ms_entry * GC_CALL GC_mark_and_push(void *obj,
+                                                mse *mark_stack_ptr,
+                                                mse *mark_stack_limit,
+                                                void **src)
+{
+    hdr * hhdr;
+
+    PREFETCH(obj);
+    GET_HDR(obj, hhdr);
+    if (EXPECT(IS_FORWARDING_ADDR_OR_NIL(hhdr), FALSE)) {
+      if (GC_all_interior_pointers) {
+        hhdr = GC_find_header(GC_base(obj));
+        if (hhdr == 0) {
+          GC_ADD_TO_BLACK_LIST_NORMAL(obj, (ptr_t)src);
+          return mark_stack_ptr;
+        }
+      } else {
+        GC_ADD_TO_BLACK_LIST_NORMAL(obj, (ptr_t)src);
+        return mark_stack_ptr;
+      }
+    }
+    if (EXPECT(HBLK_IS_FREE(hhdr), FALSE)) {
+        GC_ADD_TO_BLACK_LIST_NORMAL(obj, (ptr_t)src);
+        return mark_stack_ptr;
+    }
+
+    PUSH_CONTENTS_HDR(obj, mark_stack_ptr /* modified */, mark_stack_limit,
+                      (ptr_t)src, was_marked, hhdr, TRUE);
+ was_marked:
+    return mark_stack_ptr;
+}
+
+#if defined(MANUAL_VDB) && defined(THREADS)
+  void GC_dirty(ptr_t p);
+#endif
+
+/* Mark and push (i.e. gray) a single object p onto the main    */
+/* mark stack.  Consider p to be valid if it is an interior     */
+/* pointer.                                                     */
+/* The object p has passed a preliminary pointer validity       */
+/* test, but we do not definitely know whether it is valid.     */
+/* Mark bits are NOT atomically updated.  Thus this must be the */
+/* only thread setting them.                                    */
+# if defined(PRINT_BLACK_LIST) || defined(KEEP_BACK_PTRS)
+    GC_INNER void GC_mark_and_push_stack(ptr_t p, ptr_t source)
+# else
+    GC_INNER void GC_mark_and_push_stack(ptr_t p)
+#   define source ((ptr_t)0)
+# endif
+{
+    hdr * hhdr;
+    ptr_t r = p;
+
+    PREFETCH(p);
+    GET_HDR(p, hhdr);
+    if (EXPECT(IS_FORWARDING_ADDR_OR_NIL(hhdr), FALSE)) {
+        if (hhdr != 0) {
+          r = GC_base(p);
+          hhdr = HDR(r);
+        }
+        if (hhdr == 0) {
+            GC_ADD_TO_BLACK_LIST_STACK(p, source);
+            return;
+        }
+    }
+    if (EXPECT(HBLK_IS_FREE(hhdr), FALSE)) {
+        GC_ADD_TO_BLACK_LIST_NORMAL(p, source);
+        return;
+    }
+#   if defined(MANUAL_VDB) && defined(THREADS)
+      /* Pointer is on the stack.  We may have dirtied the object       */
+      /* it points to, but not yet have called GC_dirty();              */
+      GC_dirty(p);      /* Implicitly affects entire object.            */
+#   endif
+    PUSH_CONTENTS_HDR(r, GC_mark_stack_top, GC_mark_stack_limit,
+                      source, mark_and_push_exit, hhdr, FALSE);
+  mark_and_push_exit: ;
+    /* We silently ignore pointers to near the end of a block,  */
+    /* which is very mildly suboptimal.                         */
+    /* FIXME: We should probably add a header word to address   */
+    /* this.                                                    */
+}
+# undef source
+
+# ifdef TRACE_BUF
+
+# define TRACE_ENTRIES 1000
+
+struct trace_entry {
+    char * kind;
+    word gc_no;
+    word bytes_allocd;
+    word arg1;
+    word arg2;
+} GC_trace_buf[TRACE_ENTRIES];
+
+int GC_trace_buf_ptr = 0;
+
+void GC_add_trace_entry(char *kind, word arg1, word arg2)
+{
+    GC_trace_buf[GC_trace_buf_ptr].kind = kind;
+    GC_trace_buf[GC_trace_buf_ptr].gc_no = GC_gc_no;
+    GC_trace_buf[GC_trace_buf_ptr].bytes_allocd = GC_bytes_allocd;
+    GC_trace_buf[GC_trace_buf_ptr].arg1 = arg1 ^ 0x80000000;
+    GC_trace_buf[GC_trace_buf_ptr].arg2 = arg2 ^ 0x80000000;
+    GC_trace_buf_ptr++;
+    if (GC_trace_buf_ptr >= TRACE_ENTRIES) GC_trace_buf_ptr = 0;
+}
+
+void GC_print_trace(word gc_no, GC_bool lock)
+{
+    int i;
+    struct trace_entry *p;
+    DCL_LOCK_STATE;
+
+    if (lock) LOCK();
+    for (i = GC_trace_buf_ptr-1; i != GC_trace_buf_ptr; i--) {
+        if (i < 0) i = TRACE_ENTRIES-1;
+        p = GC_trace_buf + i;
+        if (p -> gc_no < gc_no || p -> kind == 0) return;
+        printf("Trace:%s (gc:%u,bytes:%lu) 0x%X, 0x%X\n",
+                p -> kind, (unsigned)p -> gc_no,
+                (unsigned long)p -> bytes_allocd,
+                (p -> arg1) ^ 0x80000000, (p -> arg2) ^ 0x80000000);
+    }
+    printf("Trace incomplete\n");
+    if (lock) UNLOCK();
+}
+
+# endif /* TRACE_BUF */
+
+/*
+ * A version of GC_push_all that treats all interior pointers as valid
+ * and scans the entire region immediately, in case the contents
+ * change.
+ */
+GC_INNER void GC_push_all_eager(ptr_t bottom, ptr_t top)
+{
+    word * b = (word *)(((word) bottom + ALIGNMENT-1) & ~(ALIGNMENT-1));
+    word * t = (word *)(((word) top) & ~(ALIGNMENT-1));
+    register word *p;
+    register word q;
+    register word *lim;
+    register ptr_t greatest_ha = GC_greatest_plausible_heap_addr;
+    register ptr_t least_ha = GC_least_plausible_heap_addr;
+#   define GC_greatest_plausible_heap_addr greatest_ha
+#   define GC_least_plausible_heap_addr least_ha
+
+    if (top == 0) return;
+    /* check all pointers in range and push if they appear      */
+    /* to be valid.                                             */
+      lim = t - 1 /* longword */;
+      for (p = b; p <= lim; p = (word *)(((ptr_t)p) + ALIGNMENT)) {
+        q = *p;
+        GC_PUSH_ONE_STACK(q, p);
+      }
+#   undef GC_greatest_plausible_heap_addr
+#   undef GC_least_plausible_heap_addr
+}
+
+GC_INNER void GC_push_all_stack(ptr_t bottom, ptr_t top)
+{
+# if defined(THREADS) && defined(MPROTECT_VDB)
+    GC_push_all_eager(bottom, top);
+# else
+    if (!NEED_FIXUP_POINTER && GC_all_interior_pointers) {
+      GC_push_all(bottom, top);
+    } else {
+      GC_push_all_eager(bottom, top);
+    }
+# endif
+}
+
+#if !defined(SMALL_CONFIG) && !defined(USE_MARK_BYTES) && \
+    defined(MARK_BIT_PER_GRANULE)
+# if GC_GRANULE_WORDS == 1
+#   define USE_PUSH_MARKED_ACCELERATORS
+#   define PUSH_GRANULE(q) \
+                { word qcontents = (q)[0]; \
+                  GC_PUSH_ONE_HEAP(qcontents, (q)); }
+# elif GC_GRANULE_WORDS == 2
+#   define USE_PUSH_MARKED_ACCELERATORS
+#   define PUSH_GRANULE(q) \
+                { word qcontents = (q)[0]; \
+                  GC_PUSH_ONE_HEAP(qcontents, (q)); \
+                  qcontents = (q)[1]; \
+                  GC_PUSH_ONE_HEAP(qcontents, (q)+1); }
+# elif GC_GRANULE_WORDS == 4
+#   define USE_PUSH_MARKED_ACCELERATORS
+#   define PUSH_GRANULE(q) \
+                { word qcontents = (q)[0]; \
+                  GC_PUSH_ONE_HEAP(qcontents, (q)); \
+                  qcontents = (q)[1]; \
+                  GC_PUSH_ONE_HEAP(qcontents, (q)+1); \
+                  qcontents = (q)[2]; \
+                  GC_PUSH_ONE_HEAP(qcontents, (q)+2); \
+                  qcontents = (q)[3]; \
+                  GC_PUSH_ONE_HEAP(qcontents, (q)+3); }
+# endif
+#endif /* !USE_MARK_BYTES && MARK_BIT_PER_GRANULE */
+
+#ifdef USE_PUSH_MARKED_ACCELERATORS
+/* Push all objects reachable from marked objects in the given block */
+/* containing objects of size 1 granule.                             */
+STATIC void GC_push_marked1(struct hblk *h, hdr *hhdr)
+{
+    word * mark_word_addr = &(hhdr->hb_marks[0]);
+    word *p;
+    word *plim;
+    word *q;
+    word mark_word;
+
+    /* Allow registers to be used for some frequently acccessed */
+    /* global variables.  Otherwise aliasing issues are likely  */
+    /* to prevent that.                                         */
+    ptr_t greatest_ha = GC_greatest_plausible_heap_addr;
+    ptr_t least_ha = GC_least_plausible_heap_addr;
+    mse * mark_stack_top = GC_mark_stack_top;
+    mse * mark_stack_limit = GC_mark_stack_limit;
+#   define GC_mark_stack_top mark_stack_top
+#   define GC_mark_stack_limit mark_stack_limit
+#   define GC_greatest_plausible_heap_addr greatest_ha
+#   define GC_least_plausible_heap_addr least_ha
+
+    p = (word *)(h->hb_body);
+    plim = (word *)(((word)h) + HBLKSIZE);
+
+    /* go through all words in block */
+        while (p < plim) {
+            mark_word = *mark_word_addr++;
+            q = p;
+            while(mark_word != 0) {
+              if (mark_word & 1) {
+                  PUSH_GRANULE(q);
+              }
+              q += GC_GRANULE_WORDS;
+              mark_word >>= 1;
+            }
+            p += WORDSZ*GC_GRANULE_WORDS;
+        }
+
+#   undef GC_greatest_plausible_heap_addr
+#   undef GC_least_plausible_heap_addr
+#   undef GC_mark_stack_top
+#   undef GC_mark_stack_limit
+
+    GC_mark_stack_top = mark_stack_top;
+}
+
+
+#ifndef UNALIGNED_PTRS
+
+/* Push all objects reachable from marked objects in the given block */
+/* of size 2 (granules) objects.                                     */
+STATIC void GC_push_marked2(struct hblk *h, hdr *hhdr)
+{
+    word * mark_word_addr = &(hhdr->hb_marks[0]);
+    word *p;
+    word *plim;
+    word *q;
+    word mark_word;
+
+    ptr_t greatest_ha = GC_greatest_plausible_heap_addr;
+    ptr_t least_ha = GC_least_plausible_heap_addr;
+    mse * mark_stack_top = GC_mark_stack_top;
+    mse * mark_stack_limit = GC_mark_stack_limit;
+
+#   define GC_mark_stack_top mark_stack_top
+#   define GC_mark_stack_limit mark_stack_limit
+#   define GC_greatest_plausible_heap_addr greatest_ha
+#   define GC_least_plausible_heap_addr least_ha
+
+    p = (word *)(h->hb_body);
+    plim = (word *)(((word)h) + HBLKSIZE);
+
+    /* go through all words in block */
+        while (p < plim) {
+            mark_word = *mark_word_addr++;
+            q = p;
+            while(mark_word != 0) {
+              if (mark_word & 1) {
+                  PUSH_GRANULE(q);
+                  PUSH_GRANULE(q + GC_GRANULE_WORDS);
+              }
+              q += 2 * GC_GRANULE_WORDS;
+              mark_word >>= 2;
+            }
+            p += WORDSZ*GC_GRANULE_WORDS;
+        }
+
+#   undef GC_greatest_plausible_heap_addr
+#   undef GC_least_plausible_heap_addr
+#   undef GC_mark_stack_top
+#   undef GC_mark_stack_limit
+
+    GC_mark_stack_top = mark_stack_top;
+}
+
+# if GC_GRANULE_WORDS < 4
+/* Push all objects reachable from marked objects in the given block */
+/* of size 4 (granules) objects.                                     */
+/* There is a risk of mark stack overflow here.  But we handle that. */
+/* And only unmarked objects get pushed, so it's not very likely.    */
+STATIC void GC_push_marked4(struct hblk *h, hdr *hhdr)
+{
+    word * mark_word_addr = &(hhdr->hb_marks[0]);
+    word *p;
+    word *plim;
+    word *q;
+    word mark_word;
+
+    ptr_t greatest_ha = GC_greatest_plausible_heap_addr;
+    ptr_t least_ha = GC_least_plausible_heap_addr;
+    mse * mark_stack_top = GC_mark_stack_top;
+    mse * mark_stack_limit = GC_mark_stack_limit;
+#   define GC_mark_stack_top mark_stack_top
+#   define GC_mark_stack_limit mark_stack_limit
+#   define GC_greatest_plausible_heap_addr greatest_ha
+#   define GC_least_plausible_heap_addr least_ha
+
+    p = (word *)(h->hb_body);
+    plim = (word *)(((word)h) + HBLKSIZE);
+
+    /* go through all words in block */
+        while (p < plim) {
+            mark_word = *mark_word_addr++;
+            q = p;
+            while(mark_word != 0) {
+              if (mark_word & 1) {
+                  PUSH_GRANULE(q);
+                  PUSH_GRANULE(q + GC_GRANULE_WORDS);
+                  PUSH_GRANULE(q + 2*GC_GRANULE_WORDS);
+                  PUSH_GRANULE(q + 3*GC_GRANULE_WORDS);
+              }
+              q += 4 * GC_GRANULE_WORDS;
+              mark_word >>= 4;
+            }
+            p += WORDSZ*GC_GRANULE_WORDS;
+        }
+#   undef GC_greatest_plausible_heap_addr
+#   undef GC_least_plausible_heap_addr
+#   undef GC_mark_stack_top
+#   undef GC_mark_stack_limit
+    GC_mark_stack_top = mark_stack_top;
+}
+
+#endif /* GC_GRANULE_WORDS < 4 */
+
+#endif /* UNALIGNED_PTRS */
+
+#endif /* USE_PUSH_MARKED_ACCELERATORS */
+
+/* Push all objects reachable from marked objects in the given block */
+STATIC void GC_push_marked(struct hblk *h, hdr *hhdr)
+{
+    size_t sz = hhdr -> hb_sz;
+    word descr = hhdr -> hb_descr;
+    ptr_t p;
+    word bit_no;
+    ptr_t lim;
+    mse * GC_mark_stack_top_reg;
+    mse * mark_stack_limit = GC_mark_stack_limit;
+
+    /* Some quick shortcuts: */
+        if ((0 | GC_DS_LENGTH) == descr) return;
+        if (GC_block_empty(hhdr)/* nothing marked */) return;
+    GC_n_rescuing_pages++;
+    GC_objects_are_marked = TRUE;
+    if (sz > MAXOBJBYTES) {
+        lim = h -> hb_body;
+    } else {
+        lim = (h + 1)->hb_body - sz;
+    }
+
+    switch(BYTES_TO_GRANULES(sz)) {
+#   if defined(USE_PUSH_MARKED_ACCELERATORS)
+     case 1:
+       GC_push_marked1(h, hhdr);
+       break;
+#    if !defined(UNALIGNED_PTRS)
+       case 2:
+         GC_push_marked2(h, hhdr);
+         break;
+#     if GC_GRANULE_WORDS < 4
+       case 4:
+         GC_push_marked4(h, hhdr);
+         break;
+#     endif
+#    endif
+#   endif
+     default:
+      GC_mark_stack_top_reg = GC_mark_stack_top;
+      for (p = h -> hb_body, bit_no = 0; p <= lim;
+           p += sz, bit_no += MARK_BIT_OFFSET(sz)) {
+         if (mark_bit_from_hdr(hhdr, bit_no)) {
+           /* Mark from fields inside the object */
+             PUSH_OBJ(p, hhdr, GC_mark_stack_top_reg, mark_stack_limit);
+         }
+      }
+      GC_mark_stack_top = GC_mark_stack_top_reg;
+    }
+}
+
+#ifndef GC_DISABLE_INCREMENTAL
+  /* Test whether any page in the given block is dirty.   */
+  STATIC GC_bool GC_block_was_dirty(struct hblk *h, hdr *hhdr)
+  {
+    size_t sz = hhdr -> hb_sz;
+
+    if (sz <= MAXOBJBYTES) {
+         return(GC_page_was_dirty(h));
+    } else {
+         ptr_t p = (ptr_t)h;
+         while (p < (ptr_t)h + sz) {
+             if (GC_page_was_dirty((struct hblk *)p)) return(TRUE);
+             p += HBLKSIZE;
+         }
+         return(FALSE);
+    }
+  }
+#endif /* GC_DISABLE_INCREMENTAL */
+
+/* Similar to GC_push_marked, but skip over unallocated blocks  */
+/* and return address of next plausible block.                  */
+STATIC struct hblk * GC_push_next_marked(struct hblk *h)
+{
+    hdr * hhdr = HDR(h);
+
+    if (EXPECT(IS_FORWARDING_ADDR_OR_NIL(hhdr) || HBLK_IS_FREE(hhdr), FALSE)) {
+      h = GC_next_used_block(h);
+      if (h == 0) return(0);
+      hhdr = GC_find_header((ptr_t)h);
+    }
+    GC_push_marked(h, hhdr);
+    return(h + OBJ_SZ_TO_BLOCKS(hhdr -> hb_sz));
+}
+
+#ifndef GC_DISABLE_INCREMENTAL
+  /* Identical to above, but mark only from dirty pages   */
+  STATIC struct hblk * GC_push_next_marked_dirty(struct hblk *h)
+  {
+    hdr * hhdr = HDR(h);
+
+    if (!GC_dirty_maintained) ABORT("Dirty bits not set up");
+    for (;;) {
+        if (EXPECT(IS_FORWARDING_ADDR_OR_NIL(hhdr)
+                   || HBLK_IS_FREE(hhdr), FALSE)) {
+          h = GC_next_used_block(h);
+          if (h == 0) return(0);
+          hhdr = GC_find_header((ptr_t)h);
+        }
+#       ifdef STUBBORN_ALLOC
+          if (hhdr -> hb_obj_kind == STUBBORN) {
+            if (GC_page_was_changed(h) && GC_block_was_dirty(h, hhdr)) {
+                break;
+            }
+          } else {
+            if (GC_block_was_dirty(h, hhdr)) break;
+          }
+#       else
+          if (GC_block_was_dirty(h, hhdr)) break;
+#       endif
+        h += OBJ_SZ_TO_BLOCKS(hhdr -> hb_sz);
+        hhdr = HDR(h);
+    }
+    GC_push_marked(h, hhdr);
+    return(h + OBJ_SZ_TO_BLOCKS(hhdr -> hb_sz));
+  }
+#endif /* !GC_DISABLE_INCREMENTAL */
+
+/* Similar to above, but for uncollectable pages.  Needed since we      */
+/* do not clear marks for such pages, even for full collections.        */
+STATIC struct hblk * GC_push_next_marked_uncollectable(struct hblk *h)
+{
+    hdr * hhdr = HDR(h);
+
+    for (;;) {
+        if (EXPECT(IS_FORWARDING_ADDR_OR_NIL(hhdr)
+                   || HBLK_IS_FREE(hhdr), FALSE)) {
+          h = GC_next_used_block(h);
+          if (h == 0) return(0);
+          hhdr = GC_find_header((ptr_t)h);
+        }
+        if (hhdr -> hb_obj_kind == UNCOLLECTABLE) break;
+        h += OBJ_SZ_TO_BLOCKS(hhdr -> hb_sz);
+        hhdr = HDR(h);
+    }
+    GC_push_marked(h, hhdr);
+    return(h + OBJ_SZ_TO_BLOCKS(hhdr -> hb_sz));
+}