7. Advanced topics

7.1. Finalization

In Scheme, an open file is represented by a port. In the toy Scheme interpreter, a port is a wrapper around a standard C file handle:

typedef struct port_s {
    type_t type;                    /* TYPE_PORT */
    obj_t name;                     /* name of stream */
    FILE *stream;
} port_s;

Operating systems limit the number of files that a process can have open simultaneously, so to avoid running out of file handles, it is necessary to close ports when you are done with them. If a Scheme program fails to call close-input-file, then the underlying file handle should still be closed when the port object dies. This procedure is known as finalization.

Any block in an automatically managed pool can be registered for finalization by calling mps_finalize(). In the toy Scheme interpreter, this can be done in make_port:

 static obj_t make_port(obj_t name, FILE *stream)
 {
     mps_addr_t port_ref;
     obj_t obj;
     mps_addr_t addr;
     size_t size = ALIGN_OBJ(sizeof(port_s));
     do {
         mps_res_t res = mps_reserve(&addr, obj_ap, size);
         if (res != MPS_RES_OK) error("out of memory in make_port");
         obj = addr;
         obj->port.type = TYPE_PORT;
         obj->port.name = name;
         obj->port.stream = stream;
     } while(!mps_commit(obj_ap, addr, size));
     total += sizeof(port_s);

     port_ref = obj;
     mps_finalize(arena, &port_ref);

     return obj;
 }

The MPS implements finalization by posting a message to the arena’s message queue when an object that has been registered for finalization is about to die.

If you want to finalize your objects, you must first enable finalization messages by calling mps_message_type_enable():

mps_message_type_enable(arena, mps_message_type_finalization());

You must then poll the arena’s message queue at times that are convenient for you, call mps_message_get() to pick up a finalization message from the queue, call mps_message_finalization_ref() to access the finalized object, and finally call mps_message_discard() on the finalization message. The finalized object is then subject to the normal rules of life and death: it continues to live as long as it is strongly reachable.

In the toy Scheme interpreter, the most convenient moment to process the message queue is at the start of the read–eval–print loop. When a finalization message is found, the associated file handle is closed (unless it has been closed already), and the message is discarded.

 mps_message_type_t type;

 while (mps_message_queue_type(&type, arena)) {
     mps_message_t message;
     mps_bool_t b;
     b = mps_message_get(&message, arena, type);
     assert(b); /* we just checked there was one */

     if (type == mps_message_type_finalization()) {
         mps_addr_t port_ref;
         obj_t port;
         mps_message_finalization_ref(&port_ref, arena, message);
         port = port_ref;
         assert(TYPE(port) == TYPE_PORT);
         if(port->port.stream) {
             printf("Port to file \"%s\" is dying. Closing file.\n",
                    port->port.name->string.string);
             (void)fclose(port->port.stream);
             port->port.stream = NULL;
         }
     } else {
         /* ... handle other message types ... */
     }

     mps_message_discard(arena, message);
 }

Here’s an example session showing finalization taking place:

 MPS Toy Scheme Example
 9960, 0> (open-input-file "scheme.c")
 #[port "scheme.c"]
 10064, 0> (gc)
 Collection started.
   Why: Client requests: immediate full collection.
   Clock: 3401
 Port to file "scheme.c" is dying. Closing file.
 Collection finished.
     live 10040
     condemned 10088
     not_condemned 0
     clock: 3807

It’s wise not to depend on finalization as the only method for releasing resources (see the Cautions section in Finalization), because the garbage collector does not promise to collect particular objects at particular times, and in any case it does so only when it can prove that the object is dead. So it is best to provide a reliable mechanism for releasing the resource (here, the Scheme function close-input-port), and use finalization as a backup strategy.

But this raises the possibility that a port will be closed twice: once via close-input-port and a second time via finalization. So it’s necessary to make ports robust against being closed multiple times. The toy Scheme interpreter does so by setting stream to NULL: this ensures that the file handle won’t be closed more than once.

static void port_close(obj_t port)
{
    assert(TYPE(port) == TYPE_PORT);
    if(port->port.stream != NULL) {
        fclose(port->port.stream);
        port->port.stream = NULL;
    }
}

Note that because finalization messages are processed synchronously via the message queue (and the Scheme interpreter is single-threaded) there is no need for a lock here.

It’s still possible that the toy Scheme interpreter might run out of open file handles despite having some or all of its port objects being finalizable. That’s because the arena’s message queue is only polled after evaluating an expression at top level: if the expression itself opens too many file handles, the finalization messages will queue up and not be processed in time. For example:

MPS Toy Scheme Example
9960, 0> (define (repeat n f _) (if (eqv? n 0) '() (repeat (- n 1) f (f))))
repeat
10840, 0> (repeat 300 (lambda () (open-input-file "scheme.c")) 0)
open-input-file: cannot open input file

A less naïve interpreter might process finalization messages on a more regular schedule, or might take emergency action in the event of running out of open file handles by carrying out a full garbage collection and processing any finalization messages that are posted as a result.

7.2. Location dependency

The toy Scheme interpreter contains an address-based (eq?) hash table implementation. It hashes the addresses of its keys, and so needs to take account of the possibility that a moving garbage collector might move the keys. If it fails to take account of this, the hash table might become invalid after a garbage collection.

In the interaction shown below (with a naïve version of the code) you’ll see that although the keys remain present in the table after garbage collection, they cannot be found. This is because their locations (and hence their hashes) have changed, but their positions in the table have not been updated to match.

MPS Toy Scheme Example
10240, 0> (define ht (make-eq-hashtable))
ht
10584, 0> (hashtable-set! ht 'one 1)
10768, 0> (hashtable-set! ht 'two 2)
10952, 0> (hashtable-set! ht 'three 3)
11136, 0> ht
#[hashtable (two 2) (three 3) (one 1)]
11136, 0> (hashtable-ref ht 'two #f)
2
11280, 0> (gc)
11304, 1> (hashtable-ref ht 'one #f)
#f
11448, 1> (hashtable-ref ht 'two #f)
#f
11592, 1> (hashtable-ref ht 'three #f)
#f
11736, 1> ht
#[hashtable (two 2) (three 3) (one 1)]

The MPS solves this problem with its location dependency feature: a structure of type mps_ld_s encapsulates a set of dependencies on the locations of blocks. You add addresses to the location dependency, and then later test an address to see if it is stale: that is, if the block at that address might have moved since its location was depended upon.

You need to provide space for the mps_ld_s structure. In the case of a hash table, it is most convenient to inline it in the hash table’s metadata:

typedef struct table_s {
  type_t type;                  /* TYPE_TABLE */
  hash_t hash;                  /* hash function */
  cmp_t cmp;                    /* comparison function */
  mps_ld_s ld;                  /* location dependency */
  obj_t buckets;                /* hash buckets */
} table_s;

Before being used, the location dependency must be reset to indicate that nothing is depended upon, by calling mps_ld_reset().

For example:

static obj_t make_table(size_t length, hash_t hashf, cmp_t cmpf)
{
    obj_t obj;
    mps_addr_t addr;
    size_t l, size = ALIGN_OBJ(sizeof(table_s));
    do {
        mps_res_t res = mps_reserve(&addr, obj_ap, size);
        if (res != MPS_RES_OK) error("out of memory in make_table");
        obj = addr;
        obj->table.type = TYPE_TABLE;
        obj->table.buckets = NULL;
    } while(!mps_commit(obj_ap, addr, size));
    total += size;
    obj->table.hash = hashf;
    obj->table.cmp = cmpf;
    /* round up to next power of 2 */
    for(l = 1; l < length; l *= 2);
    obj->table.buckets = make_buckets(l);
    mps_ld_reset(&obj->table.ld, arena);
    return obj;
}

Before the hash table becomes dependent on the location of a block, the address of the block must be added to its location dependency by calling mps_ld_add(). In particular, you must call mps_ld_add() before computing the hash of the address. (If you wait until afterwards, it might be too late: a garbage collection might have taken place after the hash was computed but before you added the dependency.)

In the toy Scheme interpreter, this is done just before the computation of the hash of the address.

static unsigned long eq_hash(obj_t obj, mps_ld_t ld)
{
    union {char s[sizeof(obj_t)]; obj_t addr;} u;
    if (ld) mps_ld_add(ld, arena, obj);
    u.addr = obj;
    return hash(u.s, sizeof(obj_t));
}

By adding the dependency at this point in the code, the implementation avoids adding unnecessary dependencies on a location. For example, an eqv? hash table does not need to depend on the location of numbers and characters:

static unsigned long eqv_hash(obj_t obj, mps_ld_t ld)
{
    switch(TYPE(obj)) {
        case TYPE_INTEGER:
            return obj->integer.integer;
        case TYPE_CHARACTER:
            return obj->character.c;
        default:
            return eq_hash(obj, ld);
    }
}

and a string=? hash table does not need to depend on the location of any of its keys.

Note

The garbage collector may run at any time, so a key may become stale at any time after calling mps_ld_add(), perhaps even before you’ve added it!

It’s best to postpone worrying about this until this key is actually looked up, when the staleness will be discovered. After all, it may never be looked up.

If you look up a key in an address-based hash table and fail to find it there, that might be because the table’s dependency on the location of the key is stale: that is, if the garbage collector moved the key. The function mps_ld_isstale() tells you if a block whose location you depended upon since the last call to mps_ld_reset() might have moved.

In the toy Scheme interpreter this behaviour is encapsulated into table_find:

static struct bucket_s *table_find(obj_t tbl, obj_t key, int add)
{
    struct bucket_s *b;
    assert(TYPE(tbl) == TYPE_TABLE);
    b = buckets_find(tbl, tbl->table.buckets, key, add);
    if ((b == NULL || b->key == NULL || b->key == obj_deleted)
        && mps_ld_isstale(&tbl->table.ld, arena, key))
    {
        b = table_rehash(tbl, tbl->table.buckets->buckets.length, key);
    }
    return b;
}

It’s important to test mps_ld_isstale() only in case of failure. The function may report a false positive (returning true despite the block not having moved). So if key has not moved, then if you tested mps_ld_isstale() first, it might return true and so you’d end up unnecessarily rehashing the whole table. (It’s crucial, however, to actually test that key appears in the table, not just that some key with the same hash does.)

When a table is rehashed, call mps_ld_reset() to clear the location dependency, and then mps_ld_add() for each key before it is added back to the table.

Note

After mps_ld_isstale() has returned true, and after rehashing the table, I don’t just repeat the usual lookup by calling buckets_find. That’s because the table might have become stale again already.

Instead, table_rehash finds and returns the bucket containing key. (Since it has to loop over all the entries in the table anyway, it might as well find this bucket too.)

By adding the line:

puts("stale!");

in table_find after mps_ld_isstale() returns true, it’s possible to see when the location dependency becomes stale and the table has to be rehashed:

MPS Toy Scheme Example
10240, 0> (define ht (make-eq-hashtable))
ht
10584, 0> (hashtable-set! ht 'one 1)
10768, 0> ht
#[hashtable (one 1)]
10768, 0> (gc)
10792, 1> (hashtable-ref ht 'one #f)
stale!
1
11080, 1> (hashtable-set! ht 'two 2)
11264, 1> (gc)
11288, 2> (hashtable-ref ht 'one #f)
stale!
1
11576, 2> (hashtable-set! ht 'three 3)
11760, 2> (hashtable-ref ht 'two #f)
2
11904, 2> (gc)
11928, 3> (hashtable-ref ht 'one #f)
1
12072, 3> (hashtable-ref ht 'two #f)
stale!
2
12360, 3> (hashtable-ref ht 'three #f)
3

Note

In case you’re puzzled by the highlighted lines: the symbol 'one must not have been moved by the collection, and so was found in the table at the correct location. Thus mps_ld_isstale() was not called. The symbol 'two did move in the collection, so it’s not found in the table, and that causes mps_ld_isstale() to be tested.

Don’t forget to check the location dependency for staleness when setting a value for key in a hash table, and when deleting a key from a hash table. Here’s an interaction with the toy Scheme interpreter showing a key being found to be stale when setting and when deleting it:

MPS Toy Scheme Example
13248, 0> (define ht (make-eq-hashtable))
ht
13624, 0> (hashtable-set! ht 'a 1)
13808, 0> (gc)
13832, 1> (hashtable-set! ht 'a 2)
stale!
13832, 1> (hashtable-delete! ht 'one)
stale!
14152, 1> (gc)
14176, 2> (hashtable-delete! ht 'a)
stale!
14456, 2> ht
#[hashtable]

7.3. Weak hash tables

A weak-key hash table has weak references (1) to its keys. If the key dies, the value corresponding to that key is automatically deleted from the table too. Similarly, a weak-value hash table has weak references to its values, and a doubly weak hash table has weak references to both.

In this section, I’ll describe how to add all three types of weak hash table to the toy Scheme interpreter. This requires a few far-reaching changes to the code, so in order to keep the basic integration understandable by newcomers to the MPS, I’ve made these changes in a separate version of the code:

scheme-advanced.c

The Scheme interpreter after a number of “advanced” features, including weak hash tables, have been implemented.

The MPS supports weak references only in roots and in blocks allocated in pools belonging to the AWL (Automatic Weak Linked) pool class. Roots aren’t convenient for this use case: it’s necessary for hash tables to be automatically reclaimed when they die. So AWL it is.

Note

This isn’t a design limitation of the MPS: it’s just that up until now the only uses our customers have had for weak references are the ones supported by AWL. (In particular, AWL was designed around the requirements of weak hash tables in Open Dylan.) If you need more general handling of weak references, contact us.

All the references in a formatted object belong to the same rank: that is, they are all exact, weak, or ambiguous references. In AWL, the rank of references is specified when creating an allocation point. This has consequences for the design of the hash table data structure: in weak-key strong-value hash tables, the keys need to be in one object and the values in another (and the same is true in the strong-key weak-value case). So instead of having one vector of buckets with alternate keys and values, hash tables must have two vectors, one for the keys and the other for the values, to allow keys and values to have different ranks.

These vectors will be allocated from an AWL pool with two allocation points, one for strong references, and one for weak references:

static mps_pool_t buckets_pool; /* pool for hash table buckets */
static mps_ap_t strong_buckets_ap; /* allocation point for strong buckets */
static mps_ap_t weak_buckets_ap; /* allocation point for weak buckets */

Note

It’s not necessary to allocate the strong buckets from the same pool as the weak buckets, but we’ll see below that they have to be allocated in a non-moving pool such as AWL.

The MPS splats a weak reference in a formatted object by replacing it with a null pointer when it is fixed by the object format’s scan method. So the scan method for the buckets is going to have the following structure. (See below for the actual code.)

static mps_res_t buckets_scan(mps_ss_t ss, mps_addr_t base, mps_addr_t limit)
{
    MPS_SCAN_BEGIN(ss) {
        while (base < limit) {
            buckets_t buckets = base;
            size_t length = buckets->length;
            for (i = 0; i < length; ++i) {
                mps_addr_t p = buckets->bucket[i];
                if (MPS_FIX1(ss, p)) {
                    mps_res_t res = MPS_FIX2(ss, &p);
                    if (res != MPS_RES_OK) return res;
                    if (p == NULL) {
                        /* TODO: key/value was splatted: splat value/key too */
                    }
                    buckets->bucket[i] = p;
                }
            }
            base = (char *)base +
                ALIGN_OBJ(offsetof(buckets_s, bucket) +
                          length * sizeof(buckets->bucket[0]));
        }
    } MPS_SCAN_END(ss);
    return MPS_RES_OK;
}

But how can the corresponding key/value be splatted? A format method is not normally allowed to access memory managed by the MPS in pools that might protect their objects (see the Cautions section in Object formats). The AWL pool class relaxes this constraint by allowing each object in the pool to have a dependent object. When scanning an object in an AWL pool, the MPS ensures that the dependent object is not protected. The dependent object does not have to be in the same pool as the original object, but must be in a non-moving pool. See Dependent objects.

So the value buckets will be the dependent object of the key buckets, and vice versa.

The AWL pool determines an object’s dependent object by calling a function that you supply when creating the pool. This means that each object needs to have a reference to its dependent object:

static mps_addr_t buckets_find_dependent(mps_addr_t addr)
{
    buckets_t buckets = addr;
    return buckets->dependent;
}

There’s one final requirement to take into account before revealing the new buckets structure, which is that each word in an object in an AWL pool must either be a valid word-aligned reference, or else the bottom bits of the word must be non-zero so that it does not look like an aligned pointer. So the sizes stored in the buckets structure (the length of the array of buckets, and the counts of used and deleted buckets) must be tagged so that they cannot be mistaken for pointers. See the Caution section in AWL (Automatic Weak Linked).

A one-bit tag suffices here:

#define TAG_COUNT(i) (((i) << 1) + 1)
#define UNTAG_COUNT(i) ((i) >> 1)

typedef struct buckets_s {
    struct buckets_s *dependent;  /* the dependent object */
    size_t length;                /* number of buckets (tagged) */
    size_t used;                  /* number of buckets in use (tagged) */
    size_t deleted;               /* number of deleted buckets (tagged) */
    obj_t bucket[1];              /* hash buckets */
} buckets_s, *buckets_t;

Now the full details of the scan method can be given, with the revised code highlighted:

static mps_res_t buckets_scan(mps_ss_t ss, mps_addr_t base, mps_addr_t limit)
{
    MPS_SCAN_BEGIN(ss) {
        while (base < limit) {
            buckets_t buckets = base; /* see note 1 */
            size_t i, length = UNTAG_COUNT(buckets->length);
            FIX(buckets->dependent);
            if(buckets->dependent != NULL)
                assert(buckets->dependent->length == buckets->length);
            for (i = 0; i < length; ++i) {
                mps_addr_t p = buckets->bucket[i];
                if (MPS_FIX1(ss, p)) {
                    mps_res_t res = MPS_FIX2(ss, &p);
                    if (res != MPS_RES_OK) return res;
                    if (p == NULL) {
                        /* key/value was splatted: splat value/key too */
                        p = obj_deleted; /* see note 3 */
                        buckets->deleted = TAG_COUNT(UNTAG_COUNT(buckets->deleted) + 1);
                        if (buckets->dependent != NULL) { /* see note 2 */
                            buckets->dependent->bucket[i] = p;
                            buckets->dependent->deleted
                                = TAG_COUNT(UNTAG_COUNT(buckets->dependent->deleted) + 1);
                        }
                    }
                    buckets->bucket[i] = p;
                }
            }
            base = (char *)base + ALIGN_OBJ(offsetof(buckets_s, bucket) +
                                            length * sizeof(buckets->bucket[0]));
        }
    } MPS_SCAN_END(ss);
    return MPS_RES_OK;
}

Notes

  1. There’s no need to dispatch on the type of the buckets object (or even to store a type at all) because buckets are the only objects to be stored in this pool.

  2. The dependent object must be fixed, and because the reference to it might be weak, it might be splatted. This means that even if you are confident that you will always initialize this field, you still have to guard access to it, as here.

  3. This hash table implementation uses NULL to mean “never used” and obj_deleted to mean “formerly used but then deleted”. So when a key is splatted it is necessary to replace it with obj_deleted.

The skip method is straightforward:

static mps_addr_t buckets_skip(mps_addr_t base)
{
    buckets_t buckets = base;
    size_t length = UNTAG_SIZE(buckets->length);
    return (char *)base + ALIGN_OBJ(offsetof(buckets_s, bucket) +
                                    length * sizeof(buckets->bucket[0]));
}

Now we can create the object format, the pool and the allocation points:

/* Create the buckets format. */
MPS_ARGS_BEGIN(args) {
    MPS_ARGS_ADD(args, MPS_KEY_FMT_ALIGN, ALIGNMENT);
    MPS_ARGS_ADD(args, MPS_KEY_FMT_SCAN, buckets_scan);
    MPS_ARGS_ADD(args, MPS_KEY_FMT_SKIP, buckets_skip);
    res = mps_fmt_create_k(&buckets_fmt, arena, args);
} MPS_ARGS_END(args);
if (res != MPS_RES_OK) error("Couldn't create buckets format");

/* Create an Automatic Weak Linked (AWL) pool to manage the hash table
   buckets. */
MPS_ARGS_BEGIN(args) {
    MPS_ARGS_ADD(args, MPS_KEY_FORMAT, buckets_fmt);
    MPS_ARGS_ADD(args, MPS_KEY_AWL_FIND_DEPENDENT, buckets_find_dependent);
    res = mps_pool_create_k(&buckets_pool, arena, mps_class_awl(), args);
} MPS_ARGS_END(args);
if (res != MPS_RES_OK) error("Couldn't create buckets pool");

/* Create allocation points for weak and strong buckets. */
MPS_ARGS_BEGIN(args) {
    MPS_ARGS_ADD(args, MPS_KEY_RANK, mps_rank_exact());
    res = mps_ap_create_k(&strong_buckets_ap, buckets_pool, args);
} MPS_ARGS_END(args);
if (res != MPS_RES_OK) error("Couldn't create strong buckets allocation point");
MPS_ARGS_BEGIN(args) {
    MPS_ARGS_ADD(args, MPS_KEY_RANK, mps_rank_weak());
    res = mps_ap_create_k(&weak_buckets_ap, buckets_pool, args);
} MPS_ARGS_END(args);
if (res != MPS_RES_OK) error("Couldn't create weak buckets allocation point");

By adding the line:

puts("splat!");

at the point in buckets_scan where the splatting of a weak reference is detected, we can see this happening:

MPS Toy Scheme Example
24624, 0> (define ht (make-doubly-weak-hashtable string-hash string=?))
ht
25264, 0> (hashtable-set! ht "one" 1)
25456, 0> (hashtable-set! ht "two" 2)
25648, 0> (hashtable-set! ht "three" 3)
25840, 0> ht
#[hashtable ("two" 2) ("one" 1) ("three" 3)]
25864, 0> (gc)
splat!
splat!
splat!
25912, 1> ht
#[hashtable]

7.4. Global symbol table

In the original (non-MPS) version of the toy Scheme interpreter, the global symbol table was implemented as a key-only hash table, and each symbol stored its own name.

But now that we have weak hash tables, it makes sense to re-implement the global symbol table as a strong-key weak-value hash table mapping strings to symbols. Each symbol will now contain a reference to its name as a string object, instead of containing the name itself.

Diagram: Global symbol table design (weak references shown as dashed lines).

Global symbol table design (weak references shown as dashed lines).

This design depends on the string object containing the symbol name being immutable. As it happens, all strings are immutable, because the toy Scheme interpreter doesn’t implement string-set!, but if it did then some care would need to be taken. (Either by marking these strings as immutable in some way, or by ensuring that these strings are “private”: that is, that Scheme programs never get hold of references to them.)

When there are no more strong references to a symbol:

  1. the reference to the symbol from the “values” array may be splatted;

  2. that’s detected by the buckets scan method, which deletes the corresponding entry in the “keys” array;

  3. which may in turn cause the symbol name to die, unless there are other strong references keeping it alive.

Here’s the new symbol structure:

typedef struct symbol_s {
    type_t type;                  /* TYPE_SYMBOL */
    obj_t name;                   /* its name (a string) */
} symbol_s;

and the new implementation of intern:

static obj_t intern_string(obj_t name)
{
    obj_t symbol;
    assert(TYPE(name) == TYPE_STRING);
    symbol = table_ref(symtab, name);
    if(symbol == NULL) {
        symbol = make_symbol(name);
        table_set(symtab, name, symbol);
    }
    return symbol;
}

static obj_t intern(char *string)
{
    return intern_string(make_string(strlen(string), string));
}

The symbol table now becomes a very simple root, that only has to be registered once (not every time it is rehashed, as previously):

mps_addr_t ref;
symtab = NULL;
ref = &symtab;
res = mps_root_create_table(&symtab_root, arena, mps_rank_exact(), 0,
                            ref, 1);
if(res != MPS_RES_OK) error("Couldn't register symtab root");
symtab = make_table(16, string_hash, string_equalp, 0, 1);

Note

The order of operations is important here. The global variable symtab must be registered as a root before creating the symbol table, otherwise the symbol table might be collected in the interval between creation and registration. But we must also ensure that symtab is valid (that is, scannable) before registering it (in this case, by setting it to NULL).

By printing splat! when the splatting of a weak reference is detected by the scan method, we can see when symbols are dying:

MPS Toy Scheme Example
24624, 0> (define a 1)
a
24832, 0> '(a b c d)
(a b c d)
25144, 0> (gc)
splat!
splat!
splat!

Here, the symbols b, c and d died, but a was kept alive by the reference from the environment.

7.5. Segregation of objects

When objects of different types have different properties (different sizes, lifetimes, references, layouts) it makes sense to segregate them into pools of appropriate classes. The garbage collector in the MPS is designed to work efficiently with many pools: it traces references between objects in different pools, and it coordinates the scanning of the registers and control stacks (see Thread roots).

For example, the toy Scheme interpreter has a mixture of object types, some of which contain references to other objects (for example, pairs) that must be scanned, and some of which do not (for example, strings). If the leaf objects are segregated into a pool of an appropriate class, the cost of scanning them can be avoided.

Here the appropriate class is AMCZ (Automatic Mostly-Copying Zero-rank), and the necessary code changes are straightforward. First, global variables for the new pool and its allocation point:

static mps_pool_t leaf_pool;    /* pool for leaf objects */
static mps_ap_t leaf_ap;        /* allocation point for leaf objects */

Second, the leaf objects must be allocated on leaf_ap instead of obj_ap. And third, the pool and its allocation point must be created:

/* Create an Automatic Mostly-Copying Zero-rank (AMCZ) pool to
   manage the leaf objects. */
MPS_ARGS_BEGIN(args) {
    MPS_ARGS_ADD(args, MPS_KEY_CHAIN, obj_chain);
    MPS_ARGS_ADD(args, MPS_KEY_FORMAT, obj_fmt);
    res = mps_pool_create_k(&leaf_pool, arena, mps_class_amcz(), args);
} MPS_ARGS_END(args);
if (res != MPS_RES_OK) error("Couldn't create leaf pool");

/* Create allocation point for leaf objects. */
res = mps_ap_create_k(&leaf_ap, leaf_pool, mps_args_none);
if (res != MPS_RES_OK) error("Couldn't create leaf objects allocation point");

Note that the new pool shared a generation chain with the old pool. This is important, because the leaf objects live and die along with the non-leaf objects of similar ages.

As an initial step in making this change, the new pool uses the same object format. However, we normally wouldn’t stop there: we’d take advantage of the segregation to simplify the scanning of the objects that have been left behind.