Merged Whippet into libguile/whippet

libguile/whippet/benchmarks/README.md

@@ -0,0 +1,35 @@
+# Benchmarks
+
+ - [`mt-gcbench.c`](./mt-gcbench.c): The multi-threaded [GCBench
+   benchmark](https://hboehm.info/gc/gc_bench.html).  An old but
+   standard benchmark that allocates different sizes of binary trees.
+   As parameters it takes a heap multiplier and a number of mutator
+   threads.  We analytically compute the peak amount of live data and
+   then size the GC heap as a multiplier of that size.  It has a peak
+   heap consumption of 10 MB or so per mutator thread: not very large.
+   At a 2x heap multiplier, it causes about 30 collections for the `mmc`
+   collector, and runs somewhere around 200-400 milliseconds in
+   single-threaded mode, on the machines I have in 2022.  For low thread
+   counts, the GCBench benchmark is small; but then again many Guile
+   processes also are quite short-lived, so perhaps it is useful to
+   ensure that small heaps remain lightweight.
+
+   To stress `mmc`'s handling of fragmentation, we modified this
+   benchmark to intersperse pseudorandomly-sized holes between tree
+   nodes.
+
+ - [`quads.c`](./quads.c): A synthetic benchmark that allocates quad
+   trees.  The mutator begins by allocating one long-lived tree of depth
+   N, and then allocates 13% of the heap in depth-3 trees, 20 times,
+   simulating a fixed working set and otherwise an allocation-heavy
+   workload.  By observing the times to allocate 13% of the heap in
+   garbage we can infer mutator overheads, and also note the variance
+   for the cycles in which GC hits.
+
+## License
+
+mt-gcbench.c was originally from https://hboehm.info/gc/gc_bench/, which
+has a somewhat unclear license.  I have modified GCBench significantly
+so that I can slot in different GC implementations.  Other files are
+distributed under the Whippet license; see the top-level
+[README.md](../README.md) for more.

libguile/whippet/benchmarks/ephemerons-embedder.h

@@ -0,0 +1,54 @@
+#ifndef EPHEMERONS_EMBEDDER_H
+#define EPHEMERONS_EMBEDDER_H
+
+#include <stddef.h>
+
+#include "ephemerons-types.h"
+#include "gc-ephemeron.h"
+
+struct gc_heap;
+
+#define DEFINE_METHODS(name, Name, NAME) \
+  static inline size_t name##_size(Name *obj) GC_ALWAYS_INLINE; \
+  static inline void visit_##name##_fields(Name *obj,\
+                                           void (*visit)(struct gc_edge edge, \
+                                                         struct gc_heap *heap, \
+                                                         void *visit_data), \
+                                           struct gc_heap *heap,        \
+                                           void *visit_data) GC_ALWAYS_INLINE;
+FOR_EACH_HEAP_OBJECT_KIND(DEFINE_METHODS)
+#undef DEFINE_METHODS
+
+static inline size_t small_object_size(SmallObject *obj) { return sizeof(*obj); }
+static inline size_t ephemeron_size(Ephemeron *obj) { return gc_ephemeron_size(); }
+static inline size_t box_size(Box *obj) { return sizeof(*obj); }
+
+static inline void
+visit_small_object_fields(SmallObject *obj,
+                          void (*visit)(struct gc_edge edge, struct gc_heap *heap,
+                                        void *visit_data),
+                          struct gc_heap *heap,
+                          void *visit_data) {}
+
+static inline void
+visit_ephemeron_fields(Ephemeron *ephemeron,
+                       void (*visit)(struct gc_edge edge, struct gc_heap *heap,
+                                     void *visit_data),
+
+                       struct gc_heap *heap,
+                       void *visit_data) {
+  gc_trace_ephemeron((struct gc_ephemeron*)ephemeron, visit, heap, visit_data);
+}
+
+static inline void
+visit_box_fields(Box *box,
+                 void (*visit)(struct gc_edge edge, struct gc_heap *heap,
+                               void *visit_data),
+                 struct gc_heap *heap,
+                 void *visit_data) {
+  visit(gc_edge(&box->obj), heap, visit_data);
+}
+
+#include "simple-gc-embedder.h"
+
+#endif // EPHEMERONS_EMBEDDER_H

libguile/whippet/benchmarks/ephemerons-types.h

@@ -0,0 +1,21 @@
+#ifndef EPHEMERONS_TYPES_H
+#define EPHEMERONS_TYPES_H
+
+#define FOR_EACH_HEAP_OBJECT_KIND(M) \
+  M(box, Box, BOX) \
+  M(ephemeron, Ephemeron, EPHEMERON) \
+  M(small_object, SmallObject, SMALL_OBJECT)
+
+#include "heap-objects.h"
+#include "simple-tagging-scheme.h"
+
+struct SmallObject {
+  struct gc_header header;
+};
+
+struct Box {
+  struct gc_header header;
+  void *obj;
+};
+
+#endif // EPHEMERONS_TYPES_H

libguile/whippet/benchmarks/ephemerons.c

@@ -0,0 +1,272 @@
+#include <errno.h>
+#include <pthread.h>
+#include <stdio.h>
+#include <stddef.h>
+#include <stdlib.h>
+#include <sys/time.h>
+
+#include "assert.h"
+#include "gc-api.h"
+#include "gc-basic-stats.h"
+#include "gc-ephemeron.h"
+#include "simple-roots-api.h"
+#include "ephemerons-types.h"
+#include "simple-allocator.h"
+
+typedef HANDLE_TO(SmallObject) SmallObjectHandle;
+typedef HANDLE_TO(struct gc_ephemeron) EphemeronHandle;
+typedef HANDLE_TO(Box) BoxHandle;
+
+static SmallObject* allocate_small_object(struct gc_mutator *mut) {
+  return gc_allocate_with_kind(mut, ALLOC_KIND_SMALL_OBJECT, sizeof(SmallObject));
+}
+
+static Box* allocate_box(struct gc_mutator *mut) {
+  return gc_allocate_with_kind(mut, ALLOC_KIND_BOX, sizeof(Box));
+}
+
+static struct gc_ephemeron* allocate_ephemeron(struct gc_mutator *mut) {
+  struct gc_ephemeron *ret = gc_allocate_ephemeron(mut);
+  *tag_word(gc_ref_from_heap_object(ret)) = tag_live(ALLOC_KIND_EPHEMERON);
+  return ret;
+}
+
+/* Get the current time in microseconds */
+static unsigned long current_time(void)
+{
+  struct timeval t;
+  if (gettimeofday(&t, NULL) == -1)
+    return 0;
+  return t.tv_sec * 1000 * 1000 + t.tv_usec;
+}
+
+struct thread {
+  struct gc_mutator *mut;
+  struct gc_mutator_roots roots;
+};
+
+static void print_elapsed(const char *what, unsigned long start) {
+  unsigned long end = current_time();
+  unsigned long msec = (end - start) / 1000;
+  unsigned long usec = (end - start) % 1000;
+  printf("Completed %s in %lu.%.3lu msec\n", what, msec, usec);
+}
+
+struct call_with_gc_data {
+  void* (*f)(struct thread *);
+  struct gc_heap *heap;
+};
+static void* call_with_gc_inner(struct gc_stack_addr *addr, void *arg) {
+  struct call_with_gc_data *data = arg;
+  struct gc_mutator *mut = gc_init_for_thread(addr, data->heap);
+  struct thread t = { mut, };
+  gc_mutator_set_roots(mut, &t.roots);
+  void *ret = data->f(&t);
+  gc_finish_for_thread(mut);
+  return ret;
+}
+static void* call_with_gc(void* (*f)(struct thread *),
+                          struct gc_heap *heap) {
+  struct call_with_gc_data data = { f, heap };
+  return gc_call_with_stack_addr(call_with_gc_inner, &data);
+}
+
+#define CHECK(x)                                                        \
+  do {                                                                  \
+    if (!(x)) {                                                         \
+      fprintf(stderr, "%s:%d: check failed: %s\n", __FILE__, __LINE__, #x); \
+      exit(1);                                                          \
+    }                                                                   \
+  } while (0)
+
+#define CHECK_EQ(x, y) CHECK((x) == (y))
+#define CHECK_NE(x, y) CHECK((x) != (y))
+#define CHECK_NULL(x) CHECK_EQ(x, NULL)
+#define CHECK_NOT_NULL(x) CHECK_NE(x, NULL)
+
+static size_t ephemeron_chain_length(struct gc_ephemeron **loc,
+                                     SmallObject *key) {
+  struct gc_ephemeron *head = gc_ephemeron_chain_head(loc);
+  size_t len = 0;
+  while (head) {
+    CHECK_EQ(key, (SmallObject*)gc_ref_value(gc_ephemeron_key(head)));
+    Box *value = gc_ref_heap_object(gc_ephemeron_value(head));
+    CHECK_NOT_NULL(value);
+    key = value->obj;
+    CHECK_NOT_NULL(key);
+    head = gc_ephemeron_chain_next(head);
+    len++;
+  }
+  return len;
+}
+
+static double heap_size;
+static double heap_multiplier;
+static size_t nthreads;
+
+static void cause_gc(struct gc_mutator *mut) {
+  // Doing a full collection lets us reason precisely about liveness.
+  gc_collect(mut, GC_COLLECTION_MAJOR);
+}
+
+static void make_ephemeron_chain(struct thread *t, EphemeronHandle *head,
+                                 SmallObjectHandle *head_key, size_t length) {
+  BoxHandle tail_box = { NULL };
+  PUSH_HANDLE(t, tail_box);
+
+  CHECK_NULL(HANDLE_REF(*head_key));
+  HANDLE_SET(*head_key, allocate_small_object(t->mut));
+
+  for (size_t i = 0; i < length; i++) {
+    HANDLE_SET(tail_box, allocate_box(t->mut));
+    HANDLE_REF(tail_box)->obj = HANDLE_REF(*head_key);
+    HANDLE_SET(*head_key, allocate_small_object(t->mut));
+    struct gc_ephemeron *ephemeron = allocate_ephemeron(t->mut);
+    gc_ephemeron_init(t->mut, ephemeron,
+                      gc_ref_from_heap_object(HANDLE_REF(*head_key)),
+                      gc_ref_from_heap_object(HANDLE_REF(tail_box)));
+    gc_ephemeron_chain_push(HANDLE_LOC(*head), ephemeron);
+  }
+
+  POP_HANDLE(t);
+}
+
+static void* run_one_test(struct thread *t) {
+  size_t unit_size = gc_ephemeron_size() + sizeof(Box);
+  size_t list_length = heap_size / nthreads / heap_multiplier / unit_size;
+
+  printf("Allocating ephemeron list %zu nodes long.  Total size %.3fGB.\n",
+         list_length, list_length * unit_size / 1e9);
+
+  unsigned long thread_start = current_time();
+
+  SmallObjectHandle head_key = { NULL };
+  EphemeronHandle head = { NULL };
+
+  PUSH_HANDLE(t, head_key);
+  PUSH_HANDLE(t, head);
+
+  make_ephemeron_chain(t, &head, &head_key, list_length);
+
+  size_t measured_length = ephemeron_chain_length(HANDLE_LOC(head),
+                                                  HANDLE_REF(head_key));
+  CHECK_EQ(measured_length, list_length);
+
+  cause_gc(t->mut);
+  measured_length = ephemeron_chain_length(HANDLE_LOC(head),
+                                           HANDLE_REF(head_key));
+  CHECK_EQ(measured_length, list_length);
+
+  if (!GC_CONSERVATIVE_ROOTS) {
+    HANDLE_SET(head_key, NULL);
+    cause_gc(t->mut);
+    measured_length = ephemeron_chain_length(HANDLE_LOC(head),
+                                             HANDLE_REF(head_key));
+    CHECK_EQ(measured_length, 0);
+  }
+
+  // swap head_key for a key halfway in, cause gc
+  // check length is expected half-length; warn, or error if precise
+  // clear and return
+
+  print_elapsed("thread", thread_start);
+
+  POP_HANDLE(t);
+  POP_HANDLE(t);
+
+  return NULL;
+}
+
+static void* run_one_test_in_thread(void *arg) {
+  struct gc_heap *heap = arg;
+  return call_with_gc(run_one_test, heap);
+}
+
+struct join_data { int status; pthread_t thread; };
+static void *join_thread(void *data) {
+  struct join_data *join_data = data;
+  void *ret;
+  join_data->status = pthread_join(join_data->thread, &ret);
+  return ret;
+}
+
+#define MAX_THREAD_COUNT 256
+
+int main(int argc, char *argv[]) {
+  if (argc < 4 || 5 < argc) {
+    fprintf(stderr, "usage: %s HEAP_SIZE MULTIPLIER NTHREADS [GC-OPTIONS]\n", argv[0]);
+    return 1;
+  }
+
+  heap_size = atof(argv[1]);
+  heap_multiplier = atof(argv[2]);
+  nthreads = atol(argv[3]);
+
+  if (heap_size < 8192) {
+    fprintf(stderr,
+            "Heap size should probably be at least 8192, right? '%s'\n",
+            argv[1]);
+    return 1;
+  }
+  if (!(1.0 < heap_multiplier && heap_multiplier < 100)) {
+    fprintf(stderr, "Failed to parse heap multiplier '%s'\n", argv[2]);
+    return 1;
+  }
+  if (nthreads < 1 || nthreads > MAX_THREAD_COUNT) {
+    fprintf(stderr, "Expected integer between 1 and %d for thread count, got '%s'\n",
+            (int)MAX_THREAD_COUNT, argv[2]);
+    return 1;
+  }
+
+  printf("Allocating heap of %.3fGB (%.2f multiplier of live data).\n",
+         heap_size / 1e9, heap_multiplier);
+
+  struct gc_options *options = gc_allocate_options();
+  gc_options_set_int(options, GC_OPTION_HEAP_SIZE_POLICY, GC_HEAP_SIZE_FIXED);
+  gc_options_set_size(options, GC_OPTION_HEAP_SIZE, heap_size);
+  if (argc == 5) {
+    if (!gc_options_parse_and_set_many(options, argv[4])) {
+      fprintf(stderr, "Failed to set GC options: '%s'\n", argv[4]);
+      return 1;
+    }
+  }
+
+  struct gc_heap *heap;
+  struct gc_mutator *mut;
+  struct gc_basic_stats stats;
+  if (!gc_init(options, NULL, &heap, &mut, GC_BASIC_STATS, &stats)) {
+    fprintf(stderr, "Failed to initialize GC with heap size %zu bytes\n",
+            (size_t)heap_size);
+    return 1;
+  }
+  struct thread main_thread = { mut, };
+  gc_mutator_set_roots(mut, &main_thread.roots);
+
+  pthread_t threads[MAX_THREAD_COUNT];
+  // Run one of the threads in the main thread.
+  for (size_t i = 1; i < nthreads; i++) {
+    int status = pthread_create(&threads[i], NULL, run_one_test_in_thread, heap);
+    if (status) {
+      errno = status;
+      perror("Failed to create thread");
+      return 1;
+    }
+  }
+  run_one_test(&main_thread);
+  for (size_t i = 1; i < nthreads; i++) {
+    struct join_data data = { 0, threads[i] };
+    gc_call_without_gc(mut, join_thread, &data);
+    if (data.status) {
+      errno = data.status;
+      perror("Failed to join thread");
+      return 1;
+    }
+  }
+  
+  gc_basic_stats_finish(&stats);
+  fputs("\n", stdout);
+  gc_basic_stats_print(&stats, stdout);
+
+  return 0;
+}
+

libguile/whippet/benchmarks/finalizers-embedder.h

@@ -0,0 +1,55 @@
+#ifndef FINALIZERS_EMBEDDER_H
+#define FINALIZERS_EMBEDDER_H
+
+#include <stddef.h>
+
+#include "finalizers-types.h"
+#include "gc-finalizer.h"
+
+struct gc_heap;
+
+#define DEFINE_METHODS(name, Name, NAME) \
+  static inline size_t name##_size(Name *obj) GC_ALWAYS_INLINE; \
+  static inline void visit_##name##_fields(Name *obj,\
+                                           void (*visit)(struct gc_edge edge, \
+                                                         struct gc_heap *heap, \
+                                                         void *visit_data), \
+                                           struct gc_heap *heap,        \
+                                           void *visit_data) GC_ALWAYS_INLINE;
+FOR_EACH_HEAP_OBJECT_KIND(DEFINE_METHODS)
+#undef DEFINE_METHODS
+
+static inline size_t small_object_size(SmallObject *obj) { return sizeof(*obj); }
+static inline size_t finalizer_size(Finalizer *obj) { return gc_finalizer_size(); }
+static inline size_t pair_size(Pair *obj) { return sizeof(*obj); }
+
+static inline void
+visit_small_object_fields(SmallObject *obj,
+                          void (*visit)(struct gc_edge edge, struct gc_heap *heap,
+                                        void *visit_data),
+                          struct gc_heap *heap,
+                          void *visit_data) {}
+
+static inline void
+visit_finalizer_fields(Finalizer *finalizer,
+                       void (*visit)(struct gc_edge edge, struct gc_heap *heap,
+                                     void *visit_data),
+
+                       struct gc_heap *heap,
+                       void *visit_data) {
+  gc_trace_finalizer((struct gc_finalizer*)finalizer, visit, heap, visit_data);
+}
+
+static inline void
+visit_pair_fields(Pair *pair,
+                  void (*visit)(struct gc_edge edge, struct gc_heap *heap,
+                                void *visit_data),
+                  struct gc_heap *heap,
+                  void *visit_data) {
+  visit(gc_edge(&pair->car), heap, visit_data);
+  visit(gc_edge(&pair->cdr), heap, visit_data);
+}
+
+#include "simple-gc-embedder.h"
+
+#endif // FINALIZERS_EMBEDDER_H

libguile/whippet/benchmarks/finalizers-types.h

@@ -0,0 +1,22 @@
+#ifndef FINALIZERS_TYPES_H
+#define FINALIZERS_TYPES_H
+
+#define FOR_EACH_HEAP_OBJECT_KIND(M) \
+  M(pair, Pair, PAIR) \
+  M(finalizer, Finalizer, FINALIZER) \
+  M(small_object, SmallObject, SMALL_OBJECT)
+
+#include "heap-objects.h"
+#include "simple-tagging-scheme.h"
+
+struct SmallObject {
+  struct gc_header header;
+};
+
+struct Pair {
+  struct gc_header header;
+  void *car;
+  void *cdr;
+};
+
+#endif // FINALIZERS_TYPES_H

libguile/whippet/benchmarks/finalizers.c

@@ -0,0 +1,284 @@
+#include <errno.h>
+#include <pthread.h>
+#include <stdio.h>
+#include <stddef.h>
+#include <stdlib.h>
+#include <sys/time.h>
+
+#include "assert.h"
+#include "gc-api.h"
+#include "gc-basic-stats.h"
+#include "gc-finalizer.h"
+#include "simple-roots-api.h"
+#include "finalizers-types.h"
+#include "simple-allocator.h"
+
+typedef HANDLE_TO(SmallObject) SmallObjectHandle;
+typedef HANDLE_TO(struct gc_finalizer) FinalizerHandle;
+typedef HANDLE_TO(Pair) PairHandle;
+
+static SmallObject* allocate_small_object(struct gc_mutator *mut) {
+  return gc_allocate_with_kind(mut, ALLOC_KIND_SMALL_OBJECT, sizeof(SmallObject));
+}
+
+static Pair* allocate_pair(struct gc_mutator *mut) {
+  return gc_allocate_with_kind(mut, ALLOC_KIND_PAIR, sizeof(Pair));
+}
+
+static struct gc_finalizer* allocate_finalizer(struct gc_mutator *mut) {
+  struct gc_finalizer *ret = gc_allocate_finalizer(mut);
+  *tag_word(gc_ref_from_heap_object(ret)) = tag_live(ALLOC_KIND_FINALIZER);
+  return ret;
+}
+
+/* Get the current time in microseconds */
+static unsigned long current_time(void)
+{
+  struct timeval t;
+  if (gettimeofday(&t, NULL) == -1)
+    return 0;
+  return t.tv_sec * 1000 * 1000 + t.tv_usec;
+}
+
+struct thread {
+  struct gc_mutator *mut;
+  struct gc_mutator_roots roots;
+};
+
+static void print_elapsed(const char *what, unsigned long start) {
+  unsigned long end = current_time();
+  unsigned long msec = (end - start) / 1000;
+  unsigned long usec = (end - start) % 1000;
+  printf("Completed %s in %lu.%.3lu msec\n", what, msec, usec);
+}
+
+struct call_with_gc_data {
+  void* (*f)(struct thread *);
+  struct gc_heap *heap;
+};
+static void* call_with_gc_inner(struct gc_stack_addr *addr, void *arg) {
+  struct call_with_gc_data *data = arg;
+  struct gc_mutator *mut = gc_init_for_thread(addr, data->heap);
+  struct thread t = { mut, };
+  gc_mutator_set_roots(mut, &t.roots);
+  void *ret = data->f(&t);
+  gc_finish_for_thread(mut);
+  return ret;
+}
+static void* call_with_gc(void* (*f)(struct thread *),
+                          struct gc_heap *heap) {
+  struct call_with_gc_data data = { f, heap };
+  return gc_call_with_stack_addr(call_with_gc_inner, &data);
+}
+
+#define CHECK(x)                                                        \
+  do {                                                                  \
+    if (!(x)) {                                                         \
+      fprintf(stderr, "%s:%d: check failed: %s\n", __FILE__, __LINE__, #x); \
+      exit(1);                                                          \
+    }                                                                   \
+  } while (0)
+
+#define CHECK_EQ(x, y) CHECK((x) == (y))
+#define CHECK_NE(x, y) CHECK((x) != (y))
+#define CHECK_NULL(x) CHECK_EQ(x, NULL)
+#define CHECK_NOT_NULL(x) CHECK_NE(x, NULL)
+
+static double heap_size;
+static double heap_multiplier;
+static size_t nthreads;
+
+static void cause_gc(struct gc_mutator *mut) {
+  // Doing a full collection lets us reason precisely about liveness.
+  gc_collect(mut, GC_COLLECTION_MAJOR);
+}
+
+static inline void set_car(struct gc_mutator *mut, Pair *obj, void *val) {
+  void **field = &obj->car;
+  if (val)
+    gc_write_barrier(mut, gc_ref_from_heap_object(obj), sizeof(Pair),
+                     gc_edge(field),
+                     gc_ref_from_heap_object(val));
+  *field = val;
+}
+
+static inline void set_cdr(struct gc_mutator *mut, Pair *obj, void *val) {
+  void **field = &obj->cdr;
+  if (val)
+    gc_write_barrier(mut, gc_ref_from_heap_object(obj), sizeof(Pair),
+                     gc_edge(field),
+                     gc_ref_from_heap_object(val));
+  field = val;
+}
+
+static Pair* make_finalizer_chain(struct thread *t, size_t length) {
+  PairHandle head = { NULL };
+  PairHandle tail = { NULL };
+  PUSH_HANDLE(t, head);
+  PUSH_HANDLE(t, tail);
+
+  for (size_t i = 0; i < length; i++) {
+    HANDLE_SET(tail, HANDLE_REF(head));
+    HANDLE_SET(head, allocate_pair(t->mut));
+    set_car(t->mut, HANDLE_REF(head), allocate_small_object(t->mut));
+    set_cdr(t->mut, HANDLE_REF(head), HANDLE_REF(tail));
+    struct gc_finalizer *finalizer = allocate_finalizer(t->mut);
+    gc_finalizer_attach(t->mut, finalizer, 0,
+                        gc_ref_from_heap_object(HANDLE_REF(head)),
+                        gc_ref_from_heap_object(HANDLE_REF(head)->car));
+  }
+
+  Pair *ret = HANDLE_REF(head);
+  POP_HANDLE(t);
+  POP_HANDLE(t);
+  return ret;
+}
+
+static void* run_one_test(struct thread *t) {
+  size_t unit_size = gc_finalizer_size() + sizeof(Pair);
+  size_t list_length = heap_size / nthreads / heap_multiplier / unit_size;
+  ssize_t outstanding = list_length;
+
+  printf("Allocating list %zu nodes long.  Total size %.3fGB.\n",
+         list_length, list_length * unit_size / 1e9);
+
+  unsigned long thread_start = current_time();
+
+  PairHandle chain = { NULL };
+  PUSH_HANDLE(t, chain);
+
+  HANDLE_SET(chain, make_finalizer_chain(t, list_length));
+  cause_gc(t->mut);
+
+  size_t finalized = 0;
+  for (struct gc_finalizer *f = gc_pop_finalizable(t->mut);
+       f;
+       f = gc_pop_finalizable(t->mut)) {
+    Pair* p = gc_ref_heap_object(gc_finalizer_object(f));
+    SmallObject* o = gc_ref_heap_object(gc_finalizer_closure(f));
+    CHECK_EQ(p->car, o);
+    finalized++;
+  }
+  printf("thread %p: GC before clear finalized %zu nodes.\n", t, finalized);
+  outstanding -= finalized;
+
+  HANDLE_SET(chain, NULL);
+  cause_gc(t->mut);
+
+  finalized = 0;
+  for (struct gc_finalizer *f = gc_pop_finalizable(t->mut);
+       f;
+       f = gc_pop_finalizable(t->mut)) {
+    Pair* p = gc_ref_heap_object(gc_finalizer_object(f));
+    SmallObject* o = gc_ref_heap_object(gc_finalizer_closure(f));
+    CHECK_EQ(p->car, o);
+    finalized++;
+  }
+  printf("thread %p: GC after clear finalized %zu nodes.\n", t, finalized);
+  outstanding -= finalized;
+
+  print_elapsed("thread", thread_start);
+
+  POP_HANDLE(t);
+
+  return (void*)outstanding;
+}
+
+static void* run_one_test_in_thread(void *arg) {
+  struct gc_heap *heap = arg;
+  return call_with_gc(run_one_test, heap);
+}
+
+struct join_data { int status; pthread_t thread; };
+static void *join_thread(void *data) {
+  struct join_data *join_data = data;
+  void *ret;
+  join_data->status = pthread_join(join_data->thread, &ret);
+  return ret;
+}
+
+#define MAX_THREAD_COUNT 256
+
+int main(int argc, char *argv[]) {
+  if (argc < 4 || 5 < argc) {
+    fprintf(stderr, "usage: %s HEAP_SIZE MULTIPLIER NTHREADS [GC-OPTIONS]\n", argv[0]);
+    return 1;
+  }
+
+  heap_size = atof(argv[1]);
+  heap_multiplier = atof(argv[2]);
+  nthreads = atol(argv[3]);
+
+  if (heap_size < 8192) {
+    fprintf(stderr,
+            "Heap size should probably be at least 8192, right? '%s'\n",
+            argv[1]);
+    return 1;
+  }
+  if (!(1.0 < heap_multiplier && heap_multiplier < 100)) {
+    fprintf(stderr, "Failed to parse heap multiplier '%s'\n", argv[2]);
+    return 1;
+  }
+  if (nthreads < 1 || nthreads > MAX_THREAD_COUNT) {
+    fprintf(stderr, "Expected integer between 1 and %d for thread count, got '%s'\n",
+            (int)MAX_THREAD_COUNT, argv[2]);
+    return 1;
+  }
+
+  printf("Allocating heap of %.3fGB (%.2f multiplier of live data).\n",
+         heap_size / 1e9, heap_multiplier);
+
+  struct gc_options *options = gc_allocate_options();
+  gc_options_set_int(options, GC_OPTION_HEAP_SIZE_POLICY, GC_HEAP_SIZE_FIXED);
+  gc_options_set_size(options, GC_OPTION_HEAP_SIZE, heap_size);
+  if (argc == 5) {
+    if (!gc_options_parse_and_set_many(options, argv[4])) {
+      fprintf(stderr, "Failed to set GC options: '%s'\n", argv[4]);
+      return 1;
+    }
+  }
+
+  struct gc_heap *heap;
+  struct gc_mutator *mut;
+  struct gc_basic_stats stats;
+  if (!gc_init(options, NULL, &heap, &mut, GC_BASIC_STATS, &stats)) {
+    fprintf(stderr, "Failed to initialize GC with heap size %zu bytes\n",
+            (size_t)heap_size);
+    return 1;
+  }
+  struct thread main_thread = { mut, };
+  gc_mutator_set_roots(mut, &main_thread.roots);
+
+  pthread_t threads[MAX_THREAD_COUNT];
+  // Run one of the threads in the main thread.
+  for (size_t i = 1; i < nthreads; i++) {
+    int status = pthread_create(&threads[i], NULL, run_one_test_in_thread, heap);
+    if (status) {
+      errno = status;
+      perror("Failed to create thread");
+      return 1;
+    }
+  }
+  ssize_t outstanding = (size_t)run_one_test(&main_thread);
+  for (size_t i = 1; i < nthreads; i++) {
+    struct join_data data = { 0, threads[i] };
+    void *ret = gc_call_without_gc(mut, join_thread, &data);
+    if (data.status) {
+      errno = data.status;
+      perror("Failed to join thread");
+      return 1;
+    }
+    ssize_t thread_outstanding = (ssize_t)ret;
+    outstanding += thread_outstanding;
+  }
+  
+  if (outstanding)
+    printf("\n\nWARNING: %zd nodes outstanding!!!\n\n", outstanding);
+
+  gc_basic_stats_finish(&stats);
+  fputs("\n", stdout);
+  gc_basic_stats_print(&stats, stdout);
+
+  return 0;
+}
+

libguile/whippet/benchmarks/heap-objects.h

@@ -0,0 +1,19 @@
+#ifndef HEAP_OBJECTS_H
+#define HEAP_OBJECTS_H
+
+#include "gc-inline.h"
+#include "gc-edge.h"
+
+#define DECLARE_NODE_TYPE(name, Name, NAME) \
+  struct Name;                              \
+  typedef struct Name Name;
+FOR_EACH_HEAP_OBJECT_KIND(DECLARE_NODE_TYPE)
+#undef DECLARE_NODE_TYPE
+
+#define DEFINE_ENUM(name, Name, NAME) ALLOC_KIND_##NAME,
+enum alloc_kind {
+  FOR_EACH_HEAP_OBJECT_KIND(DEFINE_ENUM)
+};
+#undef DEFINE_ENUM
+
+#endif // HEAP_OBJECTS_H

libguile/whippet/benchmarks/mt-gcbench-embedder.h

@@ -0,0 +1,54 @@
+#ifndef MT_GCBENCH_EMBEDDER_H
+#define MT_GCBENCH_EMBEDDER_H
+
+#include "gc-config.h"
+#include "mt-gcbench-types.h"
+
+struct gc_heap;
+
+#define DEFINE_METHODS(name, Name, NAME) \
+  static inline size_t name##_size(Name *obj) GC_ALWAYS_INLINE; \
+  static inline void visit_##name##_fields(Name *obj,\
+                                           void (*visit)(struct gc_edge edge, \
+                                                         struct gc_heap *heap, \
+                                                         void *visit_data), \
+                                           struct gc_heap *heap,        \
+                                           void *visit_data) GC_ALWAYS_INLINE;
+FOR_EACH_HEAP_OBJECT_KIND(DEFINE_METHODS)
+#undef DEFINE_METHODS
+
+static inline size_t node_size(Node *obj) {
+  return sizeof(Node);
+}
+static inline size_t double_array_size(DoubleArray *array) {
+  return sizeof(*array) + array->length * sizeof(double);
+}
+static inline size_t hole_size(Hole *hole) {
+  return sizeof(*hole) + hole->length * sizeof(uintptr_t);
+}
+static inline void
+visit_node_fields(Node *node,
+                  void (*visit)(struct gc_edge edge, struct gc_heap *heap,
+                                void *visit_data),
+                  struct gc_heap *heap, void *visit_data) {
+  visit(gc_edge(&node->left), heap, visit_data);
+  visit(gc_edge(&node->right), heap, visit_data);
+}
+static inline void
+visit_double_array_fields(DoubleArray *obj,
+                          void (*visit)(struct gc_edge edge,
+                                        struct gc_heap *heap, void *visit_data),
+                          struct gc_heap *heap, void *visit_data) {
+}
+static inline void
+visit_hole_fields(Hole *obj,
+                  void (*visit)(struct gc_edge edge,
+                                struct gc_heap *heap, void *visit_data),
+                  struct gc_heap *heap, void *visit_data) {
+  if (GC_PRECISE_ROOTS)
+    GC_CRASH();
+}
+
+#include "simple-gc-embedder.h"
+
+#endif // MT_GCBENCH_EMBEDDER_H

libguile/whippet/benchmarks/mt-gcbench-types.h

@@ -0,0 +1,34 @@
+#ifndef GCBENCH_TYPES_H
+#define GCBENCH_TYPES_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#define FOR_EACH_HEAP_OBJECT_KIND(M) \
+  M(node, Node, NODE) \
+  M(double_array, DoubleArray, DOUBLE_ARRAY) \
+  M(hole, Hole, HOLE)
+
+#include "heap-objects.h"
+#include "simple-tagging-scheme.h"
+
+struct Node {
+  struct gc_header header;
+  struct Node *left;
+  struct Node *right;
+  int i, j;
+};
+
+struct DoubleArray {
+  struct gc_header header;
+  size_t length;
+  double values[0];
+};
+
+struct Hole {
+  struct gc_header header;
+  size_t length;
+  uintptr_t values[0];
+};
+
+#endif // GCBENCH_TYPES_H

libguile/whippet/benchmarks/mt-gcbench.c

@@ -0,0 +1,402 @@
+// This is adapted from a benchmark written by John Ellis and Pete Kovac
+// of Post Communications.
+// It was modified by Hans Boehm of Silicon Graphics.
+// Translated to C++ 30 May 1997 by William D Clinger of Northeastern Univ.
+// Translated to C 15 March 2000 by Hans Boehm, now at HP Labs.
+//
+//      This is no substitute for real applications.  No actual application
+//      is likely to behave in exactly this way.  However, this benchmark was
+//      designed to be more representative of real applications than other
+//      Java GC benchmarks of which we are aware.
+//      It attempts to model those properties of allocation requests that
+//      are important to current GC techniques.
+//      It is designed to be used either to obtain a single overall performance
+//      number, or to give a more detailed estimate of how collector
+//      performance varies with object lifetimes.  It prints the time
+//      required to allocate and collect balanced binary trees of various
+//      sizes.  Smaller trees result in shorter object lifetimes.  Each cycle
+//      allocates roughly the same amount of memory.
+//      Two data structures are kept around during the entire process, so
+//      that the measured performance is representative of applications
+//      that maintain some live in-memory data.  One of these is a tree
+//      containing many pointers.  The other is a large array containing
+//      double precision floating point numbers.  Both should be of comparable
+//      size.
+//
+//      The results are only really meaningful together with a specification
+//      of how much memory was used.  It is possible to trade memory for
+//      better time performance.  This benchmark should be run in a 32 MB
+//      heap, though we don't currently know how to enforce that uniformly.
+//
+//      Unlike the original Ellis and Kovac benchmark, we do not attempt
+//      measure pause times.  This facility should eventually be added back
+//      in.  There are several reasons for omitting it for now.  The original
+//      implementation depended on assumptions about the thread scheduler
+//      that don't hold uniformly.  The results really measure both the
+//      scheduler and GC.  Pause time measurements tend to not fit well with
+//      current benchmark suites.  As far as we know, none of the current
+//      commercial Java implementations seriously attempt to minimize GC pause
+//      times.
+
+#include <errno.h>
+#include <pthread.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <sys/time.h>
+
+#include "assert.h"
+#include "gc-api.h"
+#include "gc-basic-stats.h"
+#include "mt-gcbench-types.h"
+#include "simple-roots-api.h"
+#include "simple-allocator.h"
+
+#define MAX_THREAD_COUNT 256
+
+static const int long_lived_tree_depth = 16; // about 4Mb
+static const int array_size = 500000; // about 4Mb
+static const int min_tree_depth = 4;
+static const int max_tree_depth = 16;
+
+typedef HANDLE_TO(Node) NodeHandle;
+typedef HANDLE_TO(DoubleArray) DoubleArrayHandle;
+
+static Node* allocate_node(struct gc_mutator *mut) {
+  // memset to 0 by the collector.
+  return gc_allocate_with_kind(mut, ALLOC_KIND_NODE, sizeof (Node));
+}
+
+static DoubleArray* allocate_double_array(struct gc_mutator *mut,
+                                                 size_t size) {
+  // May be uninitialized.
+  size_t bytes = sizeof(DoubleArray) + sizeof (double) * size;
+  DoubleArray *ret =
+    gc_allocate_pointerless_with_kind(mut, ALLOC_KIND_DOUBLE_ARRAY, bytes);
+  ret->length = size;
+  return ret;
+}
+
+static Hole* allocate_hole(struct gc_mutator *mut, size_t size) {
+  size_t bytes = sizeof(Hole) + sizeof (uintptr_t) * size;
+  Hole *ret = gc_allocate_with_kind(mut, ALLOC_KIND_HOLE, bytes);
+  ret->length = size;
+  return ret;
+}
+
+static unsigned long current_time(void) {
+  struct timeval t = { 0 };
+  gettimeofday(&t, NULL);
+  return t.tv_sec * 1000 * 1000 + t.tv_usec;
+}
+
+static double elapsed_millis(unsigned long start) {
+  return (current_time() - start) * 1e-3;
+}
+
+// Nodes used by a tree of a given size
+static int tree_size(int i) {
+  return ((1 << (i + 1)) - 1);
+}
+
+// Number of iterations to use for a given tree depth
+static int compute_num_iters(int i) {
+  return 2 * tree_size(max_tree_depth + 2) / tree_size(i);
+}
+
+// A power-law distribution.  Each integer was selected by starting at 0, taking
+// a random number in [0,1), and then accepting the integer if the random number
+// was less than 0.15, or trying again with the next integer otherwise.  Useful
+// for modelling allocation sizes or number of garbage objects to allocate
+// between live allocations.
+static const uint8_t power_law_distribution[256] = {
+  1, 15, 3, 12, 2, 8, 4, 0, 18, 7, 9, 8, 15, 2, 36, 5,
+  1, 9, 6, 11, 9, 19, 2, 0, 0, 3, 9, 6, 3, 2, 1, 1,
+  6, 1, 8, 4, 2, 0, 5, 3, 7, 0, 0, 3, 0, 4, 1, 7,
+  1, 8, 2, 2, 2, 14, 0, 7, 8, 0, 2, 1, 4, 12, 7, 5,
+  0, 3, 4, 13, 10, 2, 3, 7, 0, 8, 0, 23, 0, 16, 1, 1,
+  6, 28, 1, 18, 0, 3, 6, 5, 8, 6, 14, 5, 2, 5, 0, 11,
+  0, 18, 4, 16, 1, 4, 3, 13, 3, 23, 7, 4, 10, 5, 3, 13,
+  0, 14, 5, 5, 2, 5, 0, 16, 2, 0, 1, 1, 0, 0, 4, 2,
+  7, 7, 0, 5, 7, 2, 1, 24, 27, 3, 7, 1, 0, 8, 1, 4,
+  0, 3, 0, 7, 7, 3, 9, 2, 9, 2, 5, 10, 1, 1, 12, 6,
+  2, 9, 5, 0, 4, 6, 0, 7, 2, 1, 5, 4, 1, 0, 1, 15,
+  4, 0, 15, 4, 0, 0, 32, 18, 2, 2, 1, 7, 8, 3, 11, 1,
+  2, 7, 11, 1, 9, 1, 2, 6, 11, 17, 1, 2, 5, 1, 14, 3,
+  6, 1, 1, 15, 3, 1, 0, 6, 10, 8, 1, 3, 2, 7, 0, 1,
+  0, 11, 3, 3, 5, 8, 2, 0, 0, 7, 12, 2, 5, 20, 3, 7,
+  4, 4, 5, 22, 1, 5, 2, 7, 15, 2, 4, 6, 11, 8, 12, 1
+};
+
+static size_t power_law(size_t *counter) {
+  return power_law_distribution[(*counter)++ & 0xff];
+}
+
+struct thread {
+  struct gc_mutator *mut;
+  struct gc_mutator_roots roots;
+  size_t counter;
+};
+
+static void allocate_garbage(struct thread *t) {
+  size_t hole = power_law(&t->counter);
+  if (hole) {
+    allocate_hole(t->mut, hole);
+  }
+}
+
+static inline void set_field(struct gc_mutator *mut, Node *obj,
+                             Node **field, Node *val) {
+  gc_write_barrier(mut, gc_ref_from_heap_object(obj), sizeof(Node),
+                   gc_edge(field),
+                   gc_ref_from_heap_object(val));
+  *field = val;
+}
+
+// Build tree top down, assigning to older objects.
+static void populate(struct thread *t, int depth, Node *node) {
+  struct gc_mutator *mut = t->mut;
+  if (depth <= 0)
+    return;
+
+  NodeHandle self = { node };
+  PUSH_HANDLE(t, self);
+  allocate_garbage(t);
+  NodeHandle l = { allocate_node(mut) };
+  PUSH_HANDLE(t, l);
+  allocate_garbage(t);
+  NodeHandle r = { allocate_node(mut) };
+  PUSH_HANDLE(t, r);
+
+  set_field(mut, HANDLE_REF(self), &HANDLE_REF(self)->left, HANDLE_REF(l));
+  set_field(mut, HANDLE_REF(self), &HANDLE_REF(self)->right, HANDLE_REF(r));
+  // i is 0 because the memory is zeroed.
+  HANDLE_REF(self)->j = depth;
+
+  populate(t, depth-1, HANDLE_REF(self)->left);
+  populate(t, depth-1, HANDLE_REF(self)->right);
+
+  POP_HANDLE(t);
+  POP_HANDLE(t);
+  POP_HANDLE(t);
+}
+
+// Build tree bottom-up
+static Node* make_tree(struct thread *t, int depth) {
+  struct gc_mutator *mut = t->mut;
+  if (depth <= 0)
+    return allocate_node(mut);
+
+  NodeHandle left = { make_tree(t, depth-1) };
+  PUSH_HANDLE(t, left);
+  NodeHandle right = { make_tree(t, depth-1) };
+  PUSH_HANDLE(t, right);
+
+  allocate_garbage(t);
+  Node *result = allocate_node(mut);
+  result->left = HANDLE_REF(left);
+  result->right = HANDLE_REF(right);
+  // i is 0 because the memory is zeroed.
+  result->j = depth;
+
+  POP_HANDLE(t);
+  POP_HANDLE(t);
+
+  return result;
+}
+
+static void validate_tree(Node *tree, int depth) {
+#ifndef NDEBUG
+  GC_ASSERT_EQ(tree->i, 0);
+  GC_ASSERT_EQ(tree->j, depth);
+  if (depth == 0) {
+    GC_ASSERT(!tree->left);
+    GC_ASSERT(!tree->right);
+  } else {
+    GC_ASSERT(tree->left);
+    GC_ASSERT(tree->right);
+    validate_tree(tree->left, depth - 1);
+    validate_tree(tree->right, depth - 1);
+  }
+#endif
+}
+
+static void time_construction(struct thread *t, int depth) {
+  struct gc_mutator *mut = t->mut;
+  int num_iters = compute_num_iters(depth);
+  NodeHandle temp_tree = { NULL };
+  PUSH_HANDLE(t, temp_tree);
+
+  printf("Creating %d trees of depth %d\n", num_iters, depth);
+
+  {
+    unsigned long start = current_time();
+    for (int i = 0; i < num_iters; ++i) {
+      HANDLE_SET(temp_tree, allocate_node(mut));
+      populate(t, depth, HANDLE_REF(temp_tree));
+      validate_tree(HANDLE_REF(temp_tree), depth);
+      HANDLE_SET(temp_tree, NULL);
+    }
+    printf("\tTop down construction took %.3f msec\n",
+           elapsed_millis(start));
+  }
+
+  {
+    long start = current_time();
+    for (int i = 0; i < num_iters; ++i) {
+      HANDLE_SET(temp_tree, make_tree(t, depth));
+      validate_tree(HANDLE_REF(temp_tree), depth);
+      HANDLE_SET(temp_tree, NULL);
+    }
+    printf("\tBottom up construction took %.3f msec\n",
+           elapsed_millis(start));
+  }
+
+  POP_HANDLE(t);
+}
+
+struct call_with_gc_data {
+  void* (*f)(struct thread *);
+  struct gc_heap *heap;
+};
+static void* call_with_gc_inner(struct gc_stack_addr *addr, void *arg) {
+  struct call_with_gc_data *data = arg;
+  struct gc_mutator *mut = gc_init_for_thread(addr, data->heap);
+  struct thread t = { mut, };
+  gc_mutator_set_roots(mut, &t.roots);
+  void *ret = data->f(&t);
+  gc_finish_for_thread(mut);
+  return ret;
+}
+static void* call_with_gc(void* (*f)(struct thread *),
+                          struct gc_heap *heap) {
+  struct call_with_gc_data data = { f, heap };
+  return gc_call_with_stack_addr(call_with_gc_inner, &data);
+}
+
+static void* run_one_test(struct thread *t) {
+  NodeHandle long_lived_tree = { NULL };
+  NodeHandle temp_tree = { NULL };
+  DoubleArrayHandle array = { NULL };
+
+  PUSH_HANDLE(t, long_lived_tree);
+  PUSH_HANDLE(t, temp_tree);
+  PUSH_HANDLE(t, array);
+
+  // Create a long lived object
+  printf(" Creating a long-lived binary tree of depth %d\n",
+         long_lived_tree_depth);
+  HANDLE_SET(long_lived_tree, allocate_node(t->mut));
+  populate(t, long_lived_tree_depth, HANDLE_REF(long_lived_tree));
+
+  // Create long-lived array, filling half of it
+  printf(" Creating a long-lived array of %d doubles\n", array_size);
+  HANDLE_SET(array, allocate_double_array(t->mut, array_size));
+  for (int i = 0; i < array_size/2; ++i) {
+    HANDLE_REF(array)->values[i] = 1.0/i;
+  }
+
+  for (int d = min_tree_depth; d <= max_tree_depth; d += 2) {
+    time_construction(t, d);
+  }
+
+  validate_tree(HANDLE_REF(long_lived_tree), long_lived_tree_depth);
+
+  // Fake reference to LongLivedTree and array to keep them from being optimized
+  // away.
+  if (HANDLE_REF(long_lived_tree)->i != 0
+      || HANDLE_REF(array)->values[1000] != 1.0/1000)
+    fprintf(stderr, "Failed\n");
+
+  POP_HANDLE(t);
+  POP_HANDLE(t);
+  POP_HANDLE(t);
+  return NULL;
+}
+
+static void* run_one_test_in_thread(void *arg) {
+  struct gc_heap *heap = arg;
+  return call_with_gc(run_one_test, heap);
+}
+
+struct join_data { int status; pthread_t thread; };
+static void *join_thread(void *data) {
+  struct join_data *join_data = data;
+  void *ret;
+  join_data->status = pthread_join(join_data->thread, &ret);
+  return ret;
+}
+
+int main(int argc, char *argv[]) {
+  size_t heap_max_live =
+    tree_size(long_lived_tree_depth) * sizeof(Node) +
+    tree_size(max_tree_depth) * sizeof(Node) +
+    sizeof(DoubleArray) + sizeof(double) * array_size;
+  if (argc < 3 || argc > 4) {
+    fprintf(stderr, "usage: %s MULTIPLIER NTHREADS [GC-OPTIONS]\n", argv[0]);
+    return 1;
+  }
+
+  double multiplier = atof(argv[1]);
+  size_t nthreads = atol(argv[2]);
+
+  if (!(0.1 < multiplier && multiplier < 100)) {
+    fprintf(stderr, "Failed to parse heap multiplier '%s'\n", argv[1]);
+    return 1;
+  }
+  if (nthreads < 1 || nthreads > MAX_THREAD_COUNT) {
+    fprintf(stderr, "Expected integer between 1 and %d for thread count, got '%s'\n",
+            (int)MAX_THREAD_COUNT, argv[2]);
+    return 1;
+  }
+
+  size_t heap_size = heap_max_live * multiplier * nthreads;
+
+  struct gc_options *options = gc_allocate_options();
+  gc_options_set_int(options, GC_OPTION_HEAP_SIZE_POLICY, GC_HEAP_SIZE_FIXED);
+  gc_options_set_size(options, GC_OPTION_HEAP_SIZE, heap_size);
+  if (argc == 4) {
+    if (!gc_options_parse_and_set_many(options, argv[3])) {
+      fprintf(stderr, "Failed to set GC options: '%s'\n", argv[3]);
+      return 1;
+    }
+  }
+
+  struct gc_heap *heap;
+  struct gc_mutator *mut;
+  struct gc_basic_stats stats;
+  if (!gc_init(options, NULL, &heap, &mut, GC_BASIC_STATS, &stats)) {
+    fprintf(stderr, "Failed to initialize GC with heap size %zu bytes\n",
+            heap_size);
+    return 1;
+  }
+  struct thread main_thread = { mut, };
+  gc_mutator_set_roots(mut, &main_thread.roots);
+
+  printf("Garbage Collector Test\n");
+  printf(" Live storage will peak at %zd bytes.\n\n", heap_max_live);
+
+  pthread_t threads[MAX_THREAD_COUNT];
+  // Run one of the threads in the main thread.
+  for (size_t i = 1; i < nthreads; i++) {
+    int status = pthread_create(&threads[i], NULL, run_one_test_in_thread, heap);
+    if (status) {
+      errno = status;
+      perror("Failed to create thread");
+      return 1;
+    }
+  }
+  run_one_test(&main_thread);
+  for (size_t i = 1; i < nthreads; i++) {
+    struct join_data data = { 0, threads[i] };
+    gc_call_without_gc(mut, join_thread, &data);
+    if (data.status) {
+      errno = data.status;
+      perror("Failed to join thread");
+      return 1;
+    }
+  }
+  
+  gc_basic_stats_finish(&stats);
+  fputs("\n", stdout);
+  gc_basic_stats_print(&stats, stdout);
+}

libguile/whippet/benchmarks/quads-embedder.h

@@ -0,0 +1,37 @@
+#ifndef QUADS_EMBEDDER_H
+#define QUADS_EMBEDDER_H
+
+#include <stddef.h>
+
+#include "quads-types.h"
+
+struct gc_heap;
+
+#define DEFINE_METHODS(name, Name, NAME) \
+  static inline size_t name##_size(Name *obj) GC_ALWAYS_INLINE; \
+  static inline void visit_##name##_fields(Name *obj,\
+                                           void (*visit)(struct gc_edge edge, \
+                                                         struct gc_heap *heap, \
+                                                         void *visit_data), \
+                                           struct gc_heap *heap,        \
+                                           void *visit_data) GC_ALWAYS_INLINE;
+FOR_EACH_HEAP_OBJECT_KIND(DEFINE_METHODS)
+#undef DEFINE_METHODS
+
+static inline size_t quad_size(Quad *obj) {
+  return sizeof(Quad);
+}
+
+static inline void
+visit_quad_fields(Quad *quad,
+                  void (*visit)(struct gc_edge edge, struct gc_heap *heap,
+                                void *visit_data),
+                  struct gc_heap *heap,
+                  void *visit_data) {
+  for (size_t i = 0; i < 4; i++)
+    visit(gc_edge(&quad->kids[i]), heap, visit_data);
+}
+
+#include "simple-gc-embedder.h"
+
+#endif // QUADS_EMBEDDER_H

libguile/whippet/benchmarks/quads-types.h

@@ -0,0 +1,15 @@
+#ifndef QUADS_TYPES_H
+#define QUADS_TYPES_H
+
+#define FOR_EACH_HEAP_OBJECT_KIND(M) \
+  M(quad, Quad, QUAD)
+
+#include "heap-objects.h"
+#include "simple-tagging-scheme.h"
+
+struct Quad {
+  struct gc_header header;
+  struct Quad *kids[4];
+};
+
+#endif // QUADS_TYPES_H

libguile/whippet/benchmarks/quads.c

@@ -0,0 +1,181 @@
+#include <stdio.h>
+#include <stddef.h>
+#include <stdlib.h>
+#include <sys/time.h>
+
+#include "assert.h"
+#include "gc-api.h"
+#include "gc-basic-stats.h"
+#include "simple-roots-api.h"
+#include "quads-types.h"
+#include "simple-allocator.h"
+
+typedef HANDLE_TO(Quad) QuadHandle;
+
+static Quad* allocate_quad(struct gc_mutator *mut) {
+  // memset to 0 by the collector.
+  return gc_allocate_with_kind(mut, ALLOC_KIND_QUAD, sizeof (Quad));
+}
+
+/* Get the current time in microseconds */
+static unsigned long current_time(void)
+{
+  struct timeval t;
+  if (gettimeofday(&t, NULL) == -1)
+    return 0;
+  return t.tv_sec * 1000 * 1000 + t.tv_usec;
+}
+
+struct thread {
+  struct gc_mutator *mut;
+  struct gc_mutator_roots roots;
+  size_t counter;
+};
+
+// Build tree bottom-up
+static Quad* make_tree(struct thread *t, int depth) {
+  if (depth<=0) {
+    return allocate_quad(t->mut);
+  } else {
+    QuadHandle kids[4] = { { NULL }, };
+    for (size_t i = 0; i < 4; i++) {
+      HANDLE_SET(kids[i], make_tree(t, depth-1));
+      PUSH_HANDLE(t, kids[i]);
+    }
+
+    Quad *result = allocate_quad(t->mut);
+    for (size_t i = 0; i < 4; i++)
+      result->kids[i] = HANDLE_REF(kids[i]);
+
+    for (size_t i = 0; i < 4; i++)
+      POP_HANDLE(t);
+
+    return result;
+  }
+}
+
+static void validate_tree(Quad *tree, int depth) {
+  for (size_t i = 0; i < 4; i++) {
+    if (depth == 0) {
+      if (tree->kids[i])
+        abort();
+    } else {
+      if (!tree->kids[i])
+        abort();
+      validate_tree(tree->kids[i], depth - 1);
+    }
+  }
+}
+
+static void print_elapsed(const char *what, unsigned long start) {
+  unsigned long end = current_time();
+  unsigned long msec = (end - start) / 1000;
+  unsigned long usec = (end - start) % 1000;
+  printf("Completed %s in %lu.%.3lu msec\n", what, msec, usec);
+}
+
+static size_t parse_size(char *arg, const char *what) {
+  long val = atol(arg);
+  if (val <= 0) {
+    fprintf(stderr, "Failed to parse %s '%s'\n", what, arg);
+    exit(1);
+  }
+  return val;
+}
+
+static size_t tree_size(size_t depth) {
+  size_t nquads = 0;
+  size_t leaf_count = 1;
+  for (size_t i = 0; i <= depth; i++) {
+    if (nquads > ((size_t)-1) - leaf_count) {
+      fprintf(stderr,
+              "error: address space too small for quad tree of depth %zu\n",
+              depth);
+      exit(1);
+    }
+    nquads += leaf_count;
+    leaf_count *= 4;
+  }
+  return nquads;
+}
+
+#define MAX_THREAD_COUNT 256
+
+int main(int argc, char *argv[]) {
+  if (argc < 3 || 4 < argc) {
+    fprintf(stderr, "usage: %s DEPTH MULTIPLIER [GC-OPTIONS]\n", argv[0]);
+    return 1;
+  }
+
+  size_t depth = parse_size(argv[1], "depth");
+  double multiplier = atof(argv[2]);
+
+  if (!(1.0 < multiplier && multiplier < 100)) {
+    fprintf(stderr, "Failed to parse heap multiplier '%s'\n", argv[2]);
+    return 1;
+  }
+
+  size_t nquads = tree_size(depth);
+  size_t tree_bytes = nquads * sizeof(Quad);
+  size_t heap_size = tree_bytes * multiplier;
+
+  printf("Allocating heap of %.3fGB (%.2f multiplier of live data).\n",
+         heap_size / 1e9, multiplier);
+
+  struct gc_options *options = gc_allocate_options();
+  gc_options_set_int(options, GC_OPTION_HEAP_SIZE_POLICY, GC_HEAP_SIZE_FIXED);
+  gc_options_set_size(options, GC_OPTION_HEAP_SIZE, heap_size);
+  if (argc == 4) {
+    if (!gc_options_parse_and_set_many(options, argv[3])) {
+      fprintf(stderr, "Failed to set GC options: '%s'\n", argv[3]);
+      return 1;
+    }
+  }
+
+  struct gc_heap *heap;
+  struct gc_mutator *mut;
+  struct gc_basic_stats stats;
+  if (!gc_init(options, NULL, &heap, &mut, GC_BASIC_STATS, &stats)) {
+    fprintf(stderr, "Failed to initialize GC with heap size %zu bytes\n",
+            heap_size);
+    return 1;
+  }
+  struct thread t = { mut, };
+  gc_mutator_set_roots(mut, &t.roots);
+
+  QuadHandle quad = { NULL };
+
+  PUSH_HANDLE(&t, quad);
+
+  printf("Making quad tree of depth %zu (%zu nodes).  Total size %.3fGB.\n",
+         depth, nquads, (nquads * sizeof(Quad)) / 1e9);
+  unsigned long start = current_time();
+  HANDLE_SET(quad, make_tree(&t, depth));
+  print_elapsed("construction", start);
+
+  validate_tree(HANDLE_REF(quad), depth);
+
+  size_t garbage_step = heap_size / 7.5;
+  printf("Allocating %.3f GB of garbage, 20 times, validating live tree each time.\n",
+         garbage_step / 1e9);
+  unsigned long garbage_start = current_time();
+  for (size_t i = 0; i < 20; i++) {
+    size_t garbage_depth = 3;
+    start = current_time();
+    for (size_t i = garbage_step/(tree_size(garbage_depth)*4*sizeof(Quad*)); i; i--)
+      make_tree(&t, garbage_depth);
+    print_elapsed("allocating garbage", start);
+
+    start = current_time();
+    validate_tree(HANDLE_REF(quad), depth);
+  }
+  print_elapsed("allocation loop", garbage_start);
+
+  gc_basic_stats_finish(&stats);
+  fputs("\n", stdout);
+  gc_basic_stats_print(&stats, stdout);
+
+  POP_HANDLE(&t);
+  return 0;
+}
+

libguile/whippet/benchmarks/simple-allocator.h

@@ -0,0 +1,21 @@
+#ifndef SIMPLE_ALLOCATOR_H
+#define SIMPLE_ALLOCATOR_H
+
+#include "simple-tagging-scheme.h"
+#include "gc-api.h"
+
+static inline void*
+gc_allocate_with_kind(struct gc_mutator *mut, enum alloc_kind kind, size_t bytes) {
+  void *obj = gc_allocate(mut, bytes, GC_ALLOCATION_TAGGED);
+  *tag_word(gc_ref_from_heap_object(obj)) = tag_live(kind);
+  return obj;
+}
+
+static inline void*
+gc_allocate_pointerless_with_kind(struct gc_mutator *mut, enum alloc_kind kind, size_t bytes) {
+  void *obj = gc_allocate(mut, bytes, GC_ALLOCATION_TAGGED_POINTERLESS);
+  *tag_word(gc_ref_from_heap_object(obj)) = tag_live(kind);
+  return obj;
+}
+
+#endif // SIMPLE_ALLOCATOR_H

libguile/whippet/benchmarks/simple-gc-embedder.h

@@ -0,0 +1,183 @@
+#include <stdatomic.h>
+
+#include "simple-tagging-scheme.h"
+#include "simple-roots-types.h"
+#include "gc-config.h"
+#include "gc-embedder-api.h"
+
+#define GC_EMBEDDER_EPHEMERON_HEADER struct gc_header header;
+#define GC_EMBEDDER_FINALIZER_HEADER struct gc_header header;
+
+static inline size_t gc_finalizer_priority_count(void) { return 2; }
+
+static inline int
+gc_is_valid_conservative_ref_displacement(uintptr_t displacement) {
+#if GC_CONSERVATIVE_ROOTS || GC_CONSERVATIVE_TRACE
+  // Here is where you would allow tagged heap object references.
+  return displacement == 0;
+#else
+  // Shouldn't get here.
+  GC_CRASH();
+#endif
+}
+
+// No external objects in simple benchmarks.
+static inline int gc_extern_space_visit(struct gc_extern_space *space,
+                                        struct gc_edge edge,
+                                        struct gc_ref ref) {
+  GC_CRASH();
+}
+static inline void gc_extern_space_start_gc(struct gc_extern_space *space,
+                                            int is_minor_gc) {
+}
+static inline void gc_extern_space_finish_gc(struct gc_extern_space *space,
+                                             int is_minor_gc) {
+}
+
+static inline void gc_trace_object(struct gc_ref ref,
+                                   void (*trace_edge)(struct gc_edge edge,
+                                                      struct gc_heap *heap,
+                                                      void *trace_data),
+                                   struct gc_heap *heap,
+                                   void *trace_data,
+                                   size_t *size) {
+#if GC_CONSERVATIVE_TRACE
+  // Shouldn't get here.
+  GC_CRASH();
+#else
+  switch (tag_live_alloc_kind(*tag_word(ref))) {
+#define SCAN_OBJECT(name, Name, NAME)                                   \
+    case ALLOC_KIND_##NAME:                                             \
+      if (trace_edge)                                                   \
+        visit_##name##_fields(gc_ref_heap_object(ref), trace_edge,      \
+                              heap, trace_data);                        \
+      if (size)                                                         \
+        *size = name##_size(gc_ref_heap_object(ref));                   \
+      break;
+    FOR_EACH_HEAP_OBJECT_KIND(SCAN_OBJECT)
+#undef SCAN_OBJECT
+  default:
+    GC_CRASH();
+  }
+#endif
+}
+
+static inline void visit_roots(struct handle *roots,
+                               void (*trace_edge)(struct gc_edge edge,
+                                                  struct gc_heap *heap,
+                                                  void *trace_data),
+                               struct gc_heap *heap,
+                               void *trace_data) {
+  for (struct handle *h = roots; h; h = h->next)
+    trace_edge(gc_edge(&h->v), heap, trace_data);
+}
+
+static inline void gc_trace_mutator_roots(struct gc_mutator_roots *roots,
+                                          void (*trace_edge)(struct gc_edge edge,
+                                                             struct gc_heap *heap,
+                                                             void *trace_data),
+                                          struct gc_heap *heap,
+                                          void *trace_data) {
+  if (roots)
+    visit_roots(roots->roots, trace_edge, heap, trace_data);
+}
+
+static inline void gc_trace_heap_roots(struct gc_heap_roots *roots,
+                                       void (*trace_edge)(struct gc_edge edge,
+                                                          struct gc_heap *heap,
+                                                          void *trace_data),
+                                       struct gc_heap *heap,
+                                       void *trace_data) {
+  if (roots)
+    visit_roots(roots->roots, trace_edge, heap, trace_data);
+}
+
+static inline uintptr_t gc_object_forwarded_nonatomic(struct gc_ref ref) {
+  uintptr_t tag = *tag_word(ref);
+  return (tag & gcobj_not_forwarded_bit) ? 0 : tag;
+}
+
+static inline void gc_object_forward_nonatomic(struct gc_ref ref,
+                                               struct gc_ref new_ref) {
+  *tag_word(ref) = gc_ref_value(new_ref);
+}
+
+static inline struct gc_atomic_forward
+gc_atomic_forward_begin(struct gc_ref ref) {
+  uintptr_t tag = atomic_load_explicit(tag_word(ref), memory_order_acquire);
+  enum gc_forwarding_state state;
+  if (tag == gcobj_busy)
+    state = GC_FORWARDING_STATE_BUSY;
+  else if (tag & gcobj_not_forwarded_bit)
+    state = GC_FORWARDING_STATE_NOT_FORWARDED;
+  else
+    state = GC_FORWARDING_STATE_FORWARDED;
+  return (struct gc_atomic_forward){ ref, tag, state };
+}
+
+static inline int
+gc_atomic_forward_retry_busy(struct gc_atomic_forward *fwd) {
+  GC_ASSERT(fwd->state == GC_FORWARDING_STATE_BUSY);
+  uintptr_t tag = atomic_load_explicit(tag_word(fwd->ref),
+                                       memory_order_acquire);
+  if (tag == gcobj_busy)
+    return 0;
+  if (tag & gcobj_not_forwarded_bit) {
+    fwd->state = GC_FORWARDING_STATE_NOT_FORWARDED;
+    fwd->data = tag;
+  } else {
+    fwd->state = GC_FORWARDING_STATE_FORWARDED;
+    fwd->data = tag;
+  }
+  return 1;
+}
+  
+static inline void
+gc_atomic_forward_acquire(struct gc_atomic_forward *fwd) {
+  GC_ASSERT(fwd->state == GC_FORWARDING_STATE_NOT_FORWARDED);
+  if (atomic_compare_exchange_strong(tag_word(fwd->ref), &fwd->data,
+                                     gcobj_busy))
+    fwd->state = GC_FORWARDING_STATE_ACQUIRED;
+  else if (fwd->data == gcobj_busy)
+    fwd->state = GC_FORWARDING_STATE_BUSY;
+  else {
+    GC_ASSERT((fwd->data & gcobj_not_forwarded_bit) == 0);
+    fwd->state = GC_FORWARDING_STATE_FORWARDED;
+  }
+}
+
+static inline void
+gc_atomic_forward_abort(struct gc_atomic_forward *fwd) {
+  GC_ASSERT(fwd->state == GC_FORWARDING_STATE_ACQUIRED);
+  atomic_store_explicit(tag_word(fwd->ref), fwd->data, memory_order_release);
+  fwd->state = GC_FORWARDING_STATE_NOT_FORWARDED;
+}
+
+static inline size_t
+gc_atomic_forward_object_size(struct gc_atomic_forward *fwd) {
+  GC_ASSERT(fwd->state == GC_FORWARDING_STATE_ACQUIRED);
+  switch (tag_live_alloc_kind(fwd->data)) {
+#define OBJECT_SIZE(name, Name, NAME)                                   \
+    case ALLOC_KIND_##NAME:                                             \
+      return name##_size(gc_ref_heap_object(fwd->ref));
+    FOR_EACH_HEAP_OBJECT_KIND(OBJECT_SIZE)
+#undef OBJECT_SIZE
+  default:
+    GC_CRASH();
+  }
+}
+
+static inline void
+gc_atomic_forward_commit(struct gc_atomic_forward *fwd, struct gc_ref new_ref) {
+  GC_ASSERT(fwd->state == GC_FORWARDING_STATE_ACQUIRED);
+  *tag_word(new_ref) = fwd->data;
+  atomic_store_explicit(tag_word(fwd->ref), gc_ref_value(new_ref),
+                        memory_order_release);
+  fwd->state = GC_FORWARDING_STATE_FORWARDED;
+}
+
+static inline uintptr_t
+gc_atomic_forward_address(struct gc_atomic_forward *fwd) {
+  GC_ASSERT(fwd->state == GC_FORWARDING_STATE_FORWARDED);
+  return fwd->data;
+}

libguile/whippet/benchmarks/simple-roots-api.h

@@ -0,0 +1,26 @@
+#ifndef SIMPLE_ROOTS_API_H
+#define SIMPLE_ROOTS_API_H
+
+#include "gc-config.h"
+#include "simple-roots-types.h"
+
+#define HANDLE_TO(T) union { T* v; struct handle handle; }
+#define HANDLE_LOC(h) &(h).v
+#define HANDLE_REF(h) (h).v
+#define HANDLE_SET(h,val) do { (h).v = val; } while (0)
+#define PUSH_HANDLE(cx, h) push_handle(&(cx)->roots.roots, &h.handle)
+#define POP_HANDLE(cx) pop_handle(&(cx)->roots.roots)
+
+static inline void push_handle(struct handle **roots, struct handle *handle) {
+  if (GC_PRECISE_ROOTS) {
+    handle->next = *roots;
+    *roots = handle;
+  }
+}
+
+static inline void pop_handle(struct handle **roots) {
+  if (GC_PRECISE_ROOTS)
+    *roots = (*roots)->next;
+}
+
+#endif // SIMPLE_ROOTS_API_H

libguile/whippet/benchmarks/simple-roots-types.h

@@ -0,0 +1,17 @@
+#ifndef SIMPLE_ROOTS_TYPES_H
+#define SIMPLE_ROOTS_TYPES_H
+
+struct handle {
+  void *v;
+  struct handle *next;
+};
+
+struct gc_heap_roots {
+  struct handle *roots;
+};
+
+struct gc_mutator_roots {
+  struct handle *roots;
+};
+
+#endif // SIMPLE_ROOTS_TYPES_H

libguile/whippet/benchmarks/simple-tagging-scheme.h

@@ -0,0 +1,29 @@
+#ifndef SIMPLE_TAGGING_SCHEME_H
+#define SIMPLE_TAGGING_SCHEME_H
+
+#include <stdint.h>
+
+struct gc_header {
+  uintptr_t tag;
+};
+
+// Alloc kind is in bits 1-7, for live objects.
+static const uintptr_t gcobj_alloc_kind_mask = 0x7f;
+static const uintptr_t gcobj_alloc_kind_shift = 1;
+static const uintptr_t gcobj_forwarded_mask = 0x1;
+static const uintptr_t gcobj_not_forwarded_bit = 0x1;
+static const uintptr_t gcobj_busy = 0;
+static inline uint8_t tag_live_alloc_kind(uintptr_t tag) {
+  return (tag >> gcobj_alloc_kind_shift) & gcobj_alloc_kind_mask;
+}
+static inline uintptr_t tag_live(uint8_t alloc_kind) {
+  return ((uintptr_t)alloc_kind << gcobj_alloc_kind_shift)
+    | gcobj_not_forwarded_bit;
+}
+
+static inline uintptr_t* tag_word(struct gc_ref ref) {
+  struct gc_header *header = gc_ref_heap_object(ref);
+  return &header->tag;
+}
+
+#endif // SIMPLE_TAGGING_SCHEME_H