<HTML>
<HEAD>
<TITLE> Two-Level Tree Structure for Fast Pointer Lookup</TITLE>
<AUTHOR> Hans-J. Boehm, Silicon Graphics</author>
</HEAD>
<BODY>
<H1>Two-Level Tree Structure for Fast Pointer Lookup</h1>
<P>
The conservative garbage collector described
<A HREF="gc.html">here</a> uses a 2-level tree
data structure to aid in fast pointer identification.
This data structure is described in a bit more detail here, since
<OL>
<LI> Variations of the data structure are more generally useful.
<LI> It appears to be hard to understand by reading the code.
<LI> Some other collectors appear to use inferior data structures to
solve the same problem.
<LI> It is central to fast collector operation.
</ol>
A candidate pointer is divided into three sections, the <I>high</i>,
<I>middle</i>, and <I>low</i> bits. The exact division between these
three groups of bits is dependent on the detailed collector configuration.
<P>
The high and middle bits are used to look up an entry in the table described
here. The resulting table entry consists of either a block descriptor
(<TT>struct hblkhdr *</tt> or <TT>hdr *</tt>)
identifying the layout of objects in the block, or an indication that this
address range corresponds to the middle of a large block, together with a
hint for locating the actual block descriptor. Such a hint consist
of a displacement that can be subtracted from the middle bits of the candidate
pointer without leaving the object.
<P>
In either case, the block descriptor (<TT>struct hblkhdr</tt>)
refers to a table of object starting addresses (the <TT>hb_map</tt> field).
The starting address table is indexed by the low bits if the candidate pointer.
The resulting entry contains a displacement to the beginning of the object,
or an indication that this cannot be a valid object pointer.
(If all interior pointer are recognized, pointers into large objects
are handled specially, as appropriate.)
<H2>The Tree</h2>
<P>
The rest of this discussion focuses on the two level data structure
used to map the high and middle bits to the block descriptor.
<P>
The high bits are used as an index into the <TT>GC_top_index</tt> (really
<TT>GC_arrays._top_index</tt>) array. Each entry points to a
<TT>bottom_index</tt> data structure. This structure in turn consists
mostly of an array <TT>index</tt> indexed by the middle bits of
the candidate pointer. The <TT>index</tt> array contains the actual
<TT>hdr</tt> pointers.
<P>
Thus a pointer lookup consists primarily of a handful of memory references,
and can be quite fast:
<OL>
<LI> The appropriate <TT>bottom_index</tt> pointer is looked up in
<TT>GC_top_index</tt>, based on the high bits of the candidate pointer.
<LI> The appropriate <TT>hdr</tt> pointer is looked up in the
<TT>bottom_index</tt> structure, based on the middle bits.
<LI> The block layout map pointer is retrieved from the <TT>hdr</tt>
structure. (This memory reference is necessary since we try to share
block layout maps.)
<LI> The displacement to the beginning of the object is retrieved from the
above map.
</ol>
<P>
In order to conserve space, not all <TT>GC_top_index</tt> entries in fact
point to distinct <TT>bottom_index</tt> structures. If no address with
the corresponding high bits is part of the heap, then the entry points
to <TT>GC_all_nils</tt>, a single <TT>bottom_index</tt> structure consisting
only of NULL <TT>hdr</tt> pointers.
<P>
<TT>Bottom_index</tt> structures contain slightly more information than
just <TT>hdr</tt> pointers. The <TT>asc_link</tt> field is used to link
all <TT>bottom_index</tt> structures in ascending order for fast traversal.
This list is pointed to be <TT>GC_all_bottom_indices</tt>.
It is maintained with the aid of <TT>key</tt> field that contains the
high bits corresponding to the <TT>bottom_index</tt>.
<H2>64 bit addresses</h2>
<P>
In the case of 64 bit addresses, this picture is complicated slightly
by the fact that one of the index structures would have to be huge to
cover the entire address space with a two level tree. We deal with this
by turning <TT>GC_top_index</tt> into a chained hash table, instead of
a simple array. This adds a <TT>hash_link</tt> field to the
<TT>bottom_index</tt> structure.
<P>
The "hash function" consists of dropping the high bits. This is cheap to
compute, and guarantees that there will be no collisions if the heap
is contiguous and not excessively large.
<H2>A picture</h2>
<P>
The following is an ASCII diagram of the data structure.
This was contributed by Dave Barrett several years ago.
<PRE>
Data Structure used by GC_base in gc3.7:
21-Apr-94
63 LOG_TOP_SZ[11] LOG_BOTTOM_SZ[10] LOG_HBLKSIZE[13]
+------------------+----------------+------------------+------------------+
p:| | TL_HASH(hi) | | HBLKDISPL(p) |
+------------------+----------------+------------------+------------------+
\-----------------------HBLKPTR(p)-------------------/
\------------hi-------------------/
\______ ________/ \________ _______/ \________ _______/
V V V
| | |
GC_top_index[] | | |
--- +--------------+ | | |
^ | | | | |
| | | | | |
TOP +--------------+<--+ | |
_SZ +-<| [] | * | |
(items)| +--------------+ if 0 < bi< HBLKSIZE | |
| | | | then large object | |
| | | | starts at the bi'th | |
v | | | HBLK before p. | i |
--- | +--------------+ | (word- |
v | aligned) |
bi= |GET_BI(p){->hash_link}->key==hi | |
v | |
| (bottom_index) \ scratch_alloc'd | |
| ( struct bi ) / by get_index() | |
--- +->+--------------+ | |
^ | | | |
^ | | | |
BOTTOM | | ha=GET_HDR_ADDR(p) | |
_SZ(items)+--------------+<----------------------+ +-------+
| +--<| index[] | |
| | +--------------+ GC_obj_map: v
| | | | from / +-+-+-----+-+-+-+-+ ---
v | | | GC_add < 0| | | | | | | | ^
--- | +--------------+ _map_entry \ +-+-+-----+-+-+-+-+ |
| | asc_link | +-+-+-----+-+-+-+-+ MAXOBJSZ
| +--------------+ +-->| | | j | | | | | +1
| | key | | +-+-+-----+-+-+-+-+ |
| +--------------+ | +-+-+-----+-+-+-+-+ |
| | hash_link | | | | | | | | | | v
| +--------------+ | +-+-+-----+-+-+-+-+ ---
| | |<--MAX_OFFSET--->|
| | (bytes)
HDR(p)| GC_find_header(p) | |<--MAP_ENTRIES-->|
| \ from | =HBLKSIZE/WORDSZ
| (hdr) (struct hblkhdr) / alloc_hdr() | (1024 on Alpha)
+-->+----------------------+ | (8/16 bits each)
GET_HDR(p)| word hb_sz (words) | |
+----------------------+ |
| struct hblk *hb_next | |
+----------------------+ |
|mark_proc hb_mark_proc| |
+----------------------+ |
| char * hb_map |>-------------+
+----------------------+
| ushort hb_obj_kind |
+----------------------+
| hb_last_reclaimed |
--- +----------------------+
^ | |
MARK_BITS| hb_marks[] | *if hdr is free, hb_sz + DISCARD_WORDS
_SZ(words)| | is the size of a heap chunk (struct hblk)
v | | of at least MININCR*HBLKSIZE bytes (below),
--- +----------------------+ otherwise, size of each object in chunk.
Dynamic data structures above are interleaved throughout the heap in blocks of
size MININCR * HBLKSIZE bytes as done by gc_scratch_alloc which cannot be
freed; free lists are used (e.g. alloc_hdr). HBLK's below are collected.
(struct hblk)
--- +----------------------+ < HBLKSIZE --- --- DISCARD_
^ |garbage[DISCARD_WORDS]| aligned ^ ^ HDR_BYTES WORDS
| | | | v (bytes) (words)
| +-----hb_body----------+ < WORDSZ | --- ---
| | | aligned | ^ ^
| | Object 0 | | hb_sz |
| | | i |(word- (words)|
| | | (bytes)|aligned) v |
| + - - - - - - - - - - -+ --- | --- |
| | | ^ | ^ |
n * | | j (words) | hb_sz BODY_SZ
HBLKSIZE | Object 1 | v v | (words)
(bytes) | |--------------- v MAX_OFFSET
| + - - - - - - - - - - -+ --- (bytes)
| | | !All_INTERIOR_PTRS ^ |
| | | sets j only for hb_sz |
| | Object N | valid object offsets. | |
v | | All objects WORDSZ v v
--- +----------------------+ aligned. --- ---
DISCARD_WORDS is normally zero. Indeed the collector has not been tested
with another value in ages.
</pre>
</body>
