diff --git a/GCBench.c b/GCBench.c
index a16eb6eb1..6434fbdfb 100644
--- a/GCBench.c
+++ b/GCBench.c
@@ -42,10 +42,12 @@
 #include <stdlib.h>
 #include <sys/time.h>
 
-#ifdef GC_BDW
+#if defined(GC_BDW)
 #include "bdw.h"
 #elif defined(GC_SEMI)
 #include "semi.h"
+#elif defined(GC_MARK_SWEEP)
+#include "mark-sweep.h"
 #else
 #error unknown gc
 #endif
diff --git a/Makefile b/Makefile
index bd8e4dc8b..0067cec1a 100644
--- a/Makefile
+++ b/Makefile
@@ -1,5 +1,5 @@
 TESTS=GCBench # MT_GCBench MT_GCBench2
-COLLECTORS=bdw semi
+COLLECTORS=bdw semi mark-sweep
 
 CC=gcc
 CFLAGS=-Wall -O2 -g
@@ -14,6 +14,9 @@ bdw-%: bdw.h conservative-roots.h %.c
 semi-%: semi.h precise-roots.h %.c
 	$(CC) $(CFLAGS) -I. -DGC_SEMI -o $@ $*.c
 
+mark-sweep-%: mark-sweep.h precise-roots.h %.c
+	$(CC) $(CFLAGS) -I. -DGC_MARK_SWEEP -o $@ $*.c
+
 check: $(addprefix test-$(TARGET),$(TARGETS))
 
 test-%: $(ALL_TESTS)
diff --git a/mark-sweep.h b/mark-sweep.h
new file mode 100644
index 000000000..9a2d3b422
--- /dev/null
+++ b/mark-sweep.h
@@ -0,0 +1,523 @@
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+#include <sys/mman.h>
+#include <unistd.h>
+
+#include "precise-roots.h"
+
+#define STATIC_ASSERT_EQ(a, b) _Static_assert((a) == (b), "eq")
+
+#ifndef NDEBUG
+#define ASSERT(x) do { if (!(x)) __builtin_trap(); } while (0)
+#else
+#define ASSERT(x) do { } while (0)
+#endif
+#define ASSERT_EQ(a,b) ASSERT((a) == (b))
+
+#define GRANULE_SIZE 8
+#define GRANULE_SIZE_LOG_2 3
+#define LARGE_OBJECT_THRESHOLD 256
+#define LARGE_OBJECT_GRANULE_THRESHOLD 32
+
+STATIC_ASSERT_EQ(GRANULE_SIZE, 1 << GRANULE_SIZE_LOG_2);
+STATIC_ASSERT_EQ(LARGE_OBJECT_THRESHOLD,
+                 LARGE_OBJECT_GRANULE_THRESHOLD * GRANULE_SIZE);
+
+// There are small object pages for allocations of these sizes.
+#define FOR_EACH_SMALL_OBJECT_GRANULES(M) \
+  M(2) M(3) M(4) M(5) M(6) M(8) M(10) M(16) M(32)
+
+enum small_object_size {
+#define SMALL_OBJECT_GRANULE_SIZE(i) SMALL_OBJECT_##i,
+  FOR_EACH_SMALL_OBJECT_GRANULES(SMALL_OBJECT_GRANULE_SIZE)
+#undef SMALL_OBJECT_GRANULE_SIZE
+  SMALL_OBJECT_SIZES,
+  NOT_SMALL_OBJECT = SMALL_OBJECT_SIZES
+};
+
+static const uint8_t small_object_granule_sizes[] = 
+{
+#define SMALL_OBJECT_GRANULE_SIZE(i) i,
+  FOR_EACH_SMALL_OBJECT_GRANULES(SMALL_OBJECT_GRANULE_SIZE)
+#undef SMALL_OBJECT_GRANULE_SIZE
+};
+
+static enum small_object_size granules_to_small_object_size(unsigned granules) {
+  if (granules <= 1) return NOT_SMALL_OBJECT;
+#define TEST_GRANULE_SIZE(i) if (granules <= i) return SMALL_OBJECT_##i;
+  FOR_EACH_SMALL_OBJECT_GRANULES(TEST_GRANULE_SIZE);
+#undef TEST_GRANULE_SIZE
+  return NOT_SMALL_OBJECT;
+}
+  
+static uintptr_t align_up(uintptr_t addr, size_t align) {
+  return (addr + align - 1) & ~(align-1);
+}
+
+static inline size_t size_to_granules(size_t size) {
+  return (size + GRANULE_SIZE - 1) >> GRANULE_SIZE_LOG_2;
+}
+
+// Object kind is stored in low bits of first word of all heap objects
+// (allocated or free).
+enum gcobj_kind { GCOBJ_TINY, GCOBJ };
+
+// gcobj_kind is in the low bit of tag.
+static const uintptr_t gcobj_kind_bit = (1 << 0);
+static inline enum gcobj_kind tag_gcobj_kind(uintptr_t tag) {
+  return tag & gcobj_kind_bit;
+}
+
+// If bit 1 of a tag is set, the object is potentially live.  allocate()
+// returns objects with this flag set.  When sweep() adds an object to
+// the freelist, it gets added as dead (with this flag unset).  If
+// sweep() ever sees a dead object, then the object represents wasted
+// space in the form of fragmentation.
+static const uintptr_t gcobj_live_bit = (1 << 1);
+static inline int tag_maybe_live(uintptr_t tag) {
+  return tag & gcobj_live_bit;
+}
+
+// The mark bit is bit 2, and can only ever be set on allocated object
+// (i.e. never for objects on a free list).  It is cleared by the
+// sweeper before the next collection.
+static const uintptr_t gcobj_mark_bit = (1 << 2);
+static inline int tag_marked(uintptr_t tag) {
+  return tag & gcobj_mark_bit;
+}
+static inline void tag_set_marked(uintptr_t *tag_loc) {
+  *tag_loc |= gcobj_mark_bit;
+}
+static inline void tag_clear_marked(uintptr_t *tag_loc) {
+  *tag_loc &= ~gcobj_mark_bit;
+}
+
+// Alloc kind is in bits 3-10, for live objects.
+static const uintptr_t gcobj_alloc_kind_mask = 0xff;
+static const uintptr_t gcobj_alloc_kind_shift = 3;
+static inline uint8_t tag_live_alloc_kind(uintptr_t tag) {
+  return (tag >> gcobj_alloc_kind_shift) & gcobj_alloc_kind_mask;
+}
+
+// For free objects, bits 2 and up are free.  Non-tiny objects store the
+// object size in granules there.
+static const uintptr_t gcobj_free_granules_shift = 2;
+static inline uintptr_t tag_free_granules(uintptr_t tag) {
+  return tag >> gcobj_free_granules_shift;
+}
+
+static inline uintptr_t tag_free(enum gcobj_kind kind, size_t granules) {
+  return kind | (granules << gcobj_free_granules_shift);
+}
+static inline uintptr_t tag_live(enum gcobj_kind kind, uint8_t alloc_kind) {
+  return kind | gcobj_live_bit |
+    ((uintptr_t)alloc_kind << gcobj_alloc_kind_shift);
+}
+static inline uintptr_t tag_free_tiny(void) {
+  return tag_free(GCOBJ_TINY, 0);
+}
+
+// The gcobj_free_tiny and gcobj_free structs define the fields in free
+// tiny (1-granule), and non-tiny (2 granules and up) objects.
+struct gcobj_free_tiny {
+  // Low 2 bits of tag are GCOBJ_TINY, which is 0.  Bit 2 is live bit;
+  // never set for free objects.  Therefore for free objects, the
+  // 8-byte-aligned next pointer can alias the tag.
+  union {
+    uintptr_t tag;
+    struct gcobj_free_tiny *next;
+  };
+};
+
+// Objects from 2 granules and up.
+struct gcobj_free {
+  // For free objects, we store the granule size in the tag's payload.
+  // Next pointer only valid for objects on small freelist.
+  uintptr_t tag;
+  struct gcobj_free *next;
+};
+
+struct gcobj {
+  union {
+    uintptr_t tag;
+    struct gcobj_free_tiny free_tiny;
+    struct gcobj_free free;
+    uintptr_t words[0];
+    void *pointers[0];
+  };
+};
+
+static inline enum gcobj_kind gcobj_kind(struct gcobj *obj) {
+  return tag_gcobj_kind (obj->tag);
+}
+
+struct context {
+  // Segregated freelists of tiny and small objects.
+  struct gcobj_free_tiny *tiny_objects;
+  struct gcobj_free *small_objects[SMALL_OBJECT_SIZES];
+  // Unordered list of large objects.
+  struct gcobj_free *large_objects;
+  uintptr_t base;
+  size_t size;
+  uintptr_t sweep;
+  struct handle *roots;
+  long count;
+};
+
+static inline struct gcobj_free**
+get_small_object_freelist(struct context *cx, enum small_object_size kind) {
+  ASSERT(kind < SMALL_OBJECT_SIZES);
+  return &cx->small_objects[kind];
+}
+
+#define GC_HEADER uintptr_t _gc_header
+
+enum alloc_kind { NODE, DOUBLE_ARRAY };
+
+typedef void (*field_visitor)(struct context *, void **ref);
+
+static inline size_t node_size(void *obj) __attribute__((always_inline));
+static inline size_t double_array_size(void *obj) __attribute__((always_inline));
+static inline void visit_node_fields(struct context *cx, void *obj, field_visitor visit) __attribute__((always_inline));
+static inline void visit_double_array_fields(struct context *cx, void *obj, field_visitor visit) __attribute__((always_inline));
+
+static inline void clear_memory(uintptr_t addr, size_t size) {
+  memset((char*)addr, 0, size);
+}
+
+static void collect(struct context *cx) __attribute__((noinline));
+static void mark(struct context *cx, void *p);
+
+static inline void visit(struct context *cx, void **loc) {
+  mark(cx, *loc);
+}
+
+static void mark(struct context *cx, void *p) {
+  // A production mark implementation would use a worklist, to avoid
+  // stack overflow.  This implementation just uses the call stack.
+  struct gcobj *obj = p;
+  if (obj == NULL)
+    return;
+  if (tag_marked(obj->tag))
+    return;
+  tag_set_marked(&obj->tag);
+  switch (tag_live_alloc_kind(obj->tag)) {
+  case NODE:
+    visit_node_fields(cx, obj, visit);
+    break;
+  case DOUBLE_ARRAY:
+    visit_double_array_fields(cx, obj, visit);
+    break;
+  default:
+    abort ();
+  }
+}
+
+static void clear_freelists(struct context *cx) {
+  cx->tiny_objects = NULL;
+  for (int i = 0; i < SMALL_OBJECT_SIZES; i++)
+    cx->small_objects[i] = NULL;
+  cx->large_objects = NULL;
+}
+
+static void collect(struct context *cx) {
+  // fprintf(stderr, "start collect #%ld:\n", cx->count);
+  for (struct handle *h = cx->roots; h; h = h->next)
+    mark(cx, h->v);
+  // fprintf(stderr, "done marking\n");
+  cx->sweep = cx->base;
+  clear_freelists(cx);
+  cx->count++;
+}
+
+static void push_free_tiny(struct gcobj_free_tiny **loc,
+                           struct gcobj_free_tiny *obj) {
+  // Rely on obj->next having low bits being 0, indicating a non-live
+  // tiny object.
+  obj->next = *loc;
+  *loc = obj;
+}
+
+static void push_free(struct gcobj_free **loc, struct gcobj_free *obj,
+                      size_t granules) {
+  obj->tag = tag_free(GCOBJ, granules);
+  obj->next = *loc;
+  *loc = obj;
+}
+
+static void push_tiny(struct context *cx, void *obj) {
+  push_free_tiny(&cx->tiny_objects, obj);
+}
+
+static void push_small(struct context *cx, void *region,
+                       enum small_object_size kind, size_t region_granules) {
+  uintptr_t addr = (uintptr_t) region;
+  while (region_granules) {
+    size_t granules = small_object_granule_sizes[kind];
+    struct gcobj_free **loc = get_small_object_freelist(cx, kind);
+    while (granules <= region_granules) {
+      push_free(loc, (struct gcobj_free*) addr, granules);
+      region_granules -= granules;
+      addr += granules * GRANULE_SIZE;
+    }
+    if (region_granules == 1) {
+      // Region is actually a tiny object.
+      push_free_tiny(&cx->tiny_objects, (struct gcobj_free_tiny *)addr);
+      return;
+    }
+    // Fit any remaining granules into smaller freelists.
+    kind--;
+  }
+}
+
+static void push_large(struct context *cx, void *region, size_t granules) {
+  push_free(&cx->large_objects, region, granules);
+}
+
+static void reclaim(struct context *cx, void *obj, size_t granules) {
+  if (granules == 1) {
+    push_tiny(cx, obj);
+  } else if (granules <= LARGE_OBJECT_GRANULE_THRESHOLD) {
+    push_small(cx, obj, SMALL_OBJECT_SIZES - 1, granules);
+  } else {
+    push_large(cx, obj, granules);
+  }
+}
+
+static void split_large_object(struct context *cx,
+                               struct gcobj_free *large,
+                               size_t granules) {
+  size_t large_granules = tag_free_granules(large->tag);
+  ASSERT(large_granules >= granules);
+  if (large_granules == granules)
+    return;
+  
+  char *tail = ((char*)large) + granules * GRANULE_SIZE;
+  reclaim(cx, tail, large_granules - granules);
+}
+
+static void unlink_large_object(struct gcobj_free **prev,
+                                struct gcobj_free *large) {
+  *prev = large->next;
+}
+
+static size_t live_object_granules(struct gcobj *obj) {
+  enum gcobj_kind size_kind = tag_gcobj_kind(obj->tag);
+  if (size_kind == GCOBJ_TINY)
+    return 1;
+  size_t bytes;
+  switch (tag_live_alloc_kind (obj->tag)) {
+  case NODE:
+    bytes = node_size(obj);
+    break;
+  case DOUBLE_ARRAY:
+    bytes = double_array_size(obj);
+    break;
+  default:
+    abort ();
+  }
+  size_t granules = size_to_granules(bytes);
+  if (granules > LARGE_OBJECT_GRANULE_THRESHOLD)
+    return granules;
+  return small_object_granule_sizes[granules_to_small_object_size(granules)];
+}  
+
+static size_t free_object_granules(struct gcobj *obj) {
+  enum gcobj_kind size_kind = tag_gcobj_kind(obj->tag);
+  if (size_kind == GCOBJ_TINY)
+    return 1;
+  return tag_free_granules(obj->tag);
+}  
+
+// Sweep some heap to reclaim free space.  Return 1 if there is more
+// heap to sweep, or 0 if we reached the end.
+static int sweep(struct context *cx) {
+  // Sweep until we have reclaimed 128 granules (1024 kB), or we reach
+  // the end of the heap.
+  ssize_t to_reclaim = 128;
+  uintptr_t sweep = cx->sweep;
+  uintptr_t limit = cx->base + cx->size;
+
+  while (to_reclaim > 0 && sweep < limit) {
+    struct gcobj *obj = (struct gcobj*)sweep;
+    size_t obj_granules = tag_maybe_live(obj->tag)
+      ? live_object_granules(obj) : free_object_granules(obj);
+    sweep += obj_granules * GRANULE_SIZE;
+    if (tag_maybe_live(obj->tag) && tag_marked(obj->tag)) {
+      // Object survived collection; clear mark and continue sweeping.
+      tag_clear_marked(&obj->tag);
+    } else {
+      // Found a free object.  Combine with any following free objects.
+      // To avoid fragmentation, don't limit the amount to reclaim.
+      to_reclaim -= obj_granules;
+      while (sweep < limit) {
+        struct gcobj *next = (struct gcobj*)sweep;
+        if (tag_maybe_live(next->tag) && tag_marked(next->tag))
+          break;
+        size_t next_granules = tag_maybe_live(next->tag)
+          ? live_object_granules(next) : free_object_granules(next);
+        sweep += next_granules * GRANULE_SIZE;
+        to_reclaim -= next_granules;
+        obj_granules += next_granules;
+      }
+      memset(((char*)obj) + GRANULE_SIZE, 0, (obj_granules - 1) * GRANULE_SIZE);
+      reclaim(cx, obj, obj_granules);
+    }
+  }
+
+  cx->sweep = sweep;
+  return sweep < limit;
+}
+
+static void* allocate_large(struct context *cx, enum alloc_kind kind,
+                            size_t granules) {
+  int swept_from_beginning = 0;
+  struct gcobj_free *already_scanned = NULL;
+  while (1) {
+    do {
+      struct gcobj_free **prev = &cx->large_objects;
+      for (struct gcobj_free *large = cx->large_objects;
+           large != already_scanned;
+           prev = &large->next, large = large->next) {
+        if (tag_free_granules(large->tag) >= granules) {
+          unlink_large_object(prev, large);
+          split_large_object(cx, large, granules);
+          large->tag = tag_live(GCOBJ, kind);
+          return large;
+        }
+      }
+      already_scanned = cx->large_objects;
+    } while (sweep (cx));
+
+    // No large object, and we swept across the whole heap.  Collect.
+    if (swept_from_beginning) {
+      fprintf(stderr, "ran out of space, heap size %zu\n", cx->size);
+      abort();
+    } else {
+      collect(cx);
+      swept_from_beginning = 1;
+    }
+  }
+}
+  
+static void fill_small(struct context *cx, enum small_object_size kind) {
+  int swept_from_beginning = 0;
+  while (1) {
+    // First see if there are small objects already on the freelists
+    // that can be split.
+    for (enum small_object_size next_kind = kind;
+         next_kind < SMALL_OBJECT_SIZES;
+         next_kind++) {
+      struct gcobj_free **loc = get_small_object_freelist(cx, next_kind);
+      if (*loc) {
+        if (kind != next_kind) {
+          struct gcobj_free *ret = *loc;
+          *loc = ret->next;
+          push_small(cx, ret, kind,
+                     small_object_granule_sizes[next_kind]);
+        }
+        return;
+      }
+    }
+
+    // Otherwise if there is a large object, take and split it.
+    struct gcobj_free *large = cx->large_objects;
+    if (large) {
+      unlink_large_object(&cx->large_objects, large);
+      split_large_object(cx, large, LARGE_OBJECT_GRANULE_THRESHOLD);
+      push_small(cx, large, kind, LARGE_OBJECT_GRANULE_THRESHOLD);
+      return;
+    }
+
+    if (!sweep(cx)) {
+      if (swept_from_beginning) {
+        fprintf(stderr, "ran out of space, heap size %zu\n", cx->size);
+        abort();
+      } else {
+        collect(cx);
+        swept_from_beginning = 1;
+      }
+    }
+  }
+}
+
+static inline void* allocate_small(struct context *cx,
+                                   enum alloc_kind alloc_kind,
+                                   enum small_object_size small_kind) {
+  struct gcobj_free **loc = get_small_object_freelist(cx, small_kind);
+  if (!*loc)
+    fill_small(cx, small_kind);
+  struct gcobj_free *ret = *loc;
+  *loc = ret->next;
+  ret->tag = tag_live(GCOBJ, alloc_kind);
+  return (void *) ret;
+}
+
+static inline void fill_tiny(struct context *cx) {
+  struct gcobj_free **loc = get_small_object_freelist(cx, SMALL_OBJECT_2);
+  if (!*loc)
+    fill_small(cx, SMALL_OBJECT_2);
+  struct gcobj_free *small = *loc;
+  *loc = small->next;
+  struct gcobj_free_tiny *ret = (struct gcobj_free_tiny *)small;
+  reclaim(cx, ret, 1);
+  reclaim(cx, ret + 1, 1);
+}
+
+static inline void* allocate_tiny(struct context *cx,
+                                  enum alloc_kind alloc_kind) {
+  if (!cx->tiny_objects)
+    fill_tiny(cx);
+
+  struct gcobj_free_tiny *ret = cx->tiny_objects;
+  cx->tiny_objects = ret->next;
+  ret->tag = tag_live(GCOBJ_TINY, alloc_kind);
+  return ret;
+}
+
+static inline void* allocate(struct context *cx, enum alloc_kind kind,
+                             size_t size) {
+  size_t granules = size_to_granules(size);
+  if (granules <= 1)
+    return allocate_tiny(cx, kind);
+  if (granules <= LARGE_OBJECT_GRANULE_THRESHOLD)
+    return allocate_small(cx, kind, granules_to_small_object_size(granules));
+  return allocate_large(cx, kind, granules);
+}
+
+static inline void init_field(void **addr, void *val) {
+  *addr = val;
+}
+static inline void set_field(void **addr, void *val) {
+  *addr = val;
+}
+static inline void* get_field(void **addr) {
+  return *addr;
+}
+
+static inline void initialize_gc(struct context *cx, size_t size) {
+  size = align_up(size, getpagesize());
+
+  void *mem = mmap(NULL, size, PROT_READ|PROT_WRITE,
+                   MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
+  if (mem == MAP_FAILED) {
+    perror("mmap failed");
+    abort();
+  }
+  clear_freelists(cx);
+  cx->base = (uintptr_t) mem;
+  cx->size = size;
+  cx->sweep = cx->base + cx->size;
+  cx->roots = NULL;
+  cx->count = 0;
+  reclaim(cx, mem, size_to_granules(size));
+}
+
+static inline void print_start_gc_stats(struct context *cx) {
+}
+
+static inline void print_end_gc_stats(struct context *cx) {
+  printf("Completed %ld collections\n", cx->count);
+  printf("Heap size is %zd\n", cx->size);
+}