*
* This is a Binary Trees (bt) based matchfinder.
*
- * The data structure is a hash table where each hash bucket contains a binary
- * tree of sequences whose first 3 bytes share the same hash code. Each
+ * The main data structure is a hash table where each hash bucket contains a
+ * binary tree of sequences whose first 4 bytes share the same hash code. Each
* sequence is identified by its starting position in the input buffer. Each
* binary tree is always sorted such that each left child represents a sequence
* lexicographically lesser than its parent and each right child represents a
* sequence lexicographically greater than its parent.
*
* The algorithm processes the input buffer sequentially. At each byte
- * position, the hash code of the first 3 bytes of the sequence beginning at
+ * position, the hash code of the first 4 bytes of the sequence beginning at
* that position (the sequence being matched against) is computed. This
* identifies the hash bucket to use for that position. Then, a new binary tree
* node is created to represent the current sequence. Then, in a single tree
#include "wimlib/lz_extend.h"
#include "wimlib/lz_hash.h"
-#define BT_MATCHFINDER_HASH_ORDER 16
+#define BT_MATCHFINDER_HASH3_ORDER 15
+#define BT_MATCHFINDER_HASH3_WAYS 2
+#define BT_MATCHFINDER_HASH4_ORDER 16
/* TEMPLATED functions and structures have MF_SUFFIX appended to their name. */
#undef TEMPLATED
u32 offset;
};
-static inline u32
-bt_matchfinder_hash_3_bytes(const u8 *in_next)
-{
- return lz_hash_3_bytes(in_next, BT_MATCHFINDER_HASH_ORDER);
-}
-
#endif /* _WIMLIB_BT_MATCHFINDER_H */
struct TEMPLATED(bt_matchfinder) {
- pos_t hash_tab[1UL << BT_MATCHFINDER_HASH_ORDER];
- pos_t child_tab[];
+
+ /* The hash table for finding length 2 matches, if enabled */
+#ifdef BT_MATCHFINDER_HASH2_ORDER
+ mf_pos_t hash2_tab[1UL << BT_MATCHFINDER_HASH2_ORDER];
+#endif
+
+ /* The hash table for finding length 3 matches */
+ mf_pos_t hash3_tab[1UL << BT_MATCHFINDER_HASH3_ORDER][BT_MATCHFINDER_HASH3_WAYS];
+
+ /* The hash table which contains the roots of the binary trees for
+ * finding length 4+ matches */
+ mf_pos_t hash4_tab[1UL << BT_MATCHFINDER_HASH4_ORDER];
+
+ /* The child node references for the binary trees. The left and right
+ * children of the node for the sequence with position 'pos' are
+ * 'child_tab[pos * 2]' and 'child_tab[pos * 2 + 1]', respectively. */
+ mf_pos_t child_tab[];
};
/* Return the number of bytes that must be allocated for a 'bt_matchfinder' that
TEMPLATED(bt_matchfinder_size)(size_t max_bufsize)
{
return sizeof(struct TEMPLATED(bt_matchfinder)) +
- (2 * max_bufsize * sizeof(pos_t));
+ (2 * max_bufsize * sizeof(mf_pos_t));
}
/* Prepare the matchfinder for a new input buffer. */
memset(mf, 0, sizeof(*mf));
}
-static inline pos_t *
-TEMPLATED(bt_child)(struct TEMPLATED(bt_matchfinder) *mf, pos_t node, int offset)
+static inline mf_pos_t *
+TEMPLATED(bt_left_child)(struct TEMPLATED(bt_matchfinder) *mf, u32 node)
{
- return &mf->child_tab[(node << 1) + offset];
+ return &mf->child_tab[(node << 1) + 0];
}
-static inline pos_t *
-TEMPLATED(bt_left_child)(struct TEMPLATED(bt_matchfinder) *mf, pos_t node)
+static inline mf_pos_t *
+TEMPLATED(bt_right_child)(struct TEMPLATED(bt_matchfinder) *mf, u32 node)
{
- return TEMPLATED(bt_child)(mf, node, 0);
+ return &mf->child_tab[(node << 1) + 1];
}
-static inline pos_t *
-TEMPLATED(bt_right_child)(struct TEMPLATED(bt_matchfinder) *mf, pos_t node)
-{
- return TEMPLATED(bt_child)(mf, node, 1);
-}
+/* The minimum permissible value of 'max_len' for bt_matchfinder_get_matches()
+ * and bt_matchfinder_skip_position(). There must be sufficiently many bytes
+ * remaining to load a 32-bit integer from the *next* position. */
+#define BT_MATCHFINDER_REQUIRED_NBYTES 5
-/*
- * Retrieve a list of matches with the current position.
- *
- * @mf
- * The matchfinder structure.
- * @in_begin
- * Pointer to the beginning of the input buffer.
- * @in_next
- * Pointer to the next byte in the input buffer to process. This is the
- * pointer to the sequence being matched against.
- * @min_len
- * Only record matches that are at least this long.
- * @max_len
- * The maximum permissible match length at this position.
- * @nice_len
- * Stop searching if a match of at least this length is found.
- * Must be <= @max_len.
- * @max_search_depth
- * Limit on the number of potential matches to consider. Must be >= 1.
- * @next_hash
- * Pointer to the hash code for the current sequence, which was computed
- * one position in advance so that the binary tree root could be
- * prefetched. This is an input/output parameter.
- * @best_len_ret
- * The length of the longest match found is written here. (This is
- * actually redundant with the 'struct lz_match' array, but this is easier
- * for the compiler to optimize when inlined and the caller immediately
- * does a check against 'best_len'.)
- * @lz_matchptr
- * An array in which this function will record the matches. The recorded
- * matches will be sorted by strictly increasing length and strictly
- * increasing offset. The maximum number of matches that may be found is
- * 'min(nice_len, max_len) - 3 + 1'.
- *
- * The return value is a pointer to the next available slot in the @lz_matchptr
- * array. (If no matches were found, this will be the same as @lz_matchptr.)
- */
+/* Advance the binary tree matchfinder by one byte, optionally recording
+ * matches. @record_matches should be a compile-time constant. */
static inline struct lz_match *
-TEMPLATED(bt_matchfinder_get_matches)(struct TEMPLATED(bt_matchfinder) * const restrict mf,
- const u8 * const in_begin,
- const u8 * const in_next,
- const unsigned min_len,
- const unsigned max_len,
- const unsigned nice_len,
- const unsigned max_search_depth,
- u32 * restrict next_hash,
- unsigned * restrict best_len_ret,
- struct lz_match * restrict lz_matchptr)
+TEMPLATED(bt_matchfinder_advance_one_byte)(struct TEMPLATED(bt_matchfinder) * const restrict mf,
+ const u8 * const restrict in_begin,
+ const ptrdiff_t cur_pos,
+ const u32 max_len,
+ const u32 nice_len,
+ const u32 max_search_depth,
+ u32 next_hashes[const restrict static 2],
+ u32 * const restrict best_len_ret,
+ struct lz_match * restrict lz_matchptr,
+ const bool record_matches)
{
- unsigned depth_remaining = max_search_depth;
- u32 hash;
- pos_t cur_node;
+ const u8 *in_next = in_begin + cur_pos;
+ u32 depth_remaining = max_search_depth;
+ u32 next_seq4;
+ u32 next_seq3;
+ u32 hash3;
+ u32 hash4;
+#ifdef BT_MATCHFINDER_HASH2_ORDER
+ u16 seq2;
+ u32 hash2;
+#endif
+ STATIC_ASSERT(BT_MATCHFINDER_HASH3_WAYS >= 1 &&
+ BT_MATCHFINDER_HASH3_WAYS <= 2);
+ u32 cur_node;
+#if BT_MATCHFINDER_HASH3_WAYS >= 2
+ u32 cur_node_2;
+#endif
const u8 *matchptr;
- pos_t *pending_lt_ptr, *pending_gt_ptr;
- unsigned best_lt_len, best_gt_len;
- unsigned len;
- unsigned best_len = min_len - 1;
-
- if (unlikely(max_len < LZ_HASH3_REQUIRED_NBYTES + 1)) {
- *best_len_ret = best_len;
- return lz_matchptr;
+ mf_pos_t *pending_lt_ptr, *pending_gt_ptr;
+ u32 best_lt_len, best_gt_len;
+ u32 len;
+ u32 best_len = 3;
+
+ next_seq4 = load_u32_unaligned(in_next + 1);
+ next_seq3 = loaded_u32_to_u24(next_seq4);
+
+ hash3 = next_hashes[0];
+ hash4 = next_hashes[1];
+
+ next_hashes[0] = lz_hash(next_seq3, BT_MATCHFINDER_HASH3_ORDER);
+ next_hashes[1] = lz_hash(next_seq4, BT_MATCHFINDER_HASH4_ORDER);
+ prefetchw(&mf->hash3_tab[next_hashes[0]]);
+ prefetchw(&mf->hash4_tab[next_hashes[1]]);
+
+#ifdef BT_MATCHFINDER_HASH2_ORDER
+ seq2 = load_u16_unaligned(in_next);
+ hash2 = lz_hash(seq2, BT_MATCHFINDER_HASH2_ORDER);
+ cur_node = mf->hash2_tab[hash2];
+ mf->hash2_tab[hash2] = cur_pos;
+ if (record_matches &&
+ seq2 == load_u16_unaligned(&in_begin[cur_node]) &&
+ likely(in_next != in_begin))
+ {
+ lz_matchptr->length = 2;
+ lz_matchptr->offset = in_next - &in_begin[cur_node];
+ lz_matchptr++;
+ }
+#endif
+
+ cur_node = mf->hash3_tab[hash3][0];
+ mf->hash3_tab[hash3][0] = cur_pos;
+#if BT_MATCHFINDER_HASH3_WAYS >= 2
+ cur_node_2 = mf->hash3_tab[hash3][1];
+ mf->hash3_tab[hash3][1] = cur_node;
+#endif
+ if (record_matches && likely(in_next != in_begin)) {
+ u32 seq3 = load_u24_unaligned(in_next);
+ if (seq3 == load_u24_unaligned(&in_begin[cur_node])) {
+ lz_matchptr->length = 3;
+ lz_matchptr->offset = in_next - &in_begin[cur_node];
+ lz_matchptr++;
+ }
+ #if BT_MATCHFINDER_HASH3_WAYS >= 2
+ else if (seq3 == load_u24_unaligned(&in_begin[cur_node_2])) {
+ lz_matchptr->length = 3;
+ lz_matchptr->offset = in_next - &in_begin[cur_node_2];
+ lz_matchptr++;
+ }
+ #endif
}
- hash = *next_hash;
- *next_hash = bt_matchfinder_hash_3_bytes(in_next + 1);
- cur_node = mf->hash_tab[hash];
- mf->hash_tab[hash] = in_next - in_begin;
- prefetchw(&mf->hash_tab[*next_hash]);
+ cur_node = mf->hash4_tab[hash4];
+ mf->hash4_tab[hash4] = cur_pos;
- pending_lt_ptr = TEMPLATED(bt_left_child)(mf, in_next - in_begin);
- pending_gt_ptr = TEMPLATED(bt_right_child)(mf, in_next - in_begin);
- best_lt_len = 0;
- best_gt_len = 0;
- len = 0;
+ pending_lt_ptr = TEMPLATED(bt_left_child)(mf, cur_pos);
+ pending_gt_ptr = TEMPLATED(bt_right_child)(mf, cur_pos);
if (!cur_node) {
*pending_lt_ptr = 0;
return lz_matchptr;
}
+ best_lt_len = 0;
+ best_gt_len = 0;
+ len = 0;
+
for (;;) {
matchptr = &in_begin[cur_node];
if (matchptr[len] == in_next[len]) {
- len = lz_extend(in_next, matchptr, len + 1, max_len);
- if (len > best_len) {
- best_len = len;
- lz_matchptr->length = len;
- lz_matchptr->offset = in_next - matchptr;
- lz_matchptr++;
+ len = lz_extend(in_next, matchptr, len + 1,
+ (record_matches ? max_len : nice_len));
+ if (!record_matches || len > best_len) {
+ if (record_matches) {
+ best_len = len;
+ lz_matchptr->length = len;
+ lz_matchptr->offset = in_next - matchptr;
+ lz_matchptr++;
+ }
if (len >= nice_len) {
*pending_lt_ptr = *TEMPLATED(bt_left_child)(mf, cur_node);
*pending_gt_ptr = *TEMPLATED(bt_right_child)(mf, cur_node);
}
/*
- * Advance the matchfinder, but don't record any matches.
+ * Retrieve a list of matches with the current position.
*
* @mf
* The matchfinder structure.
* @in_begin
* Pointer to the beginning of the input buffer.
- * @in_next
- * Pointer to the next byte in the input buffer to process.
- * @in_end
- * Pointer to the end of the input buffer.
+ * @cur_pos
+ * The current position in the input buffer (the position of the sequence
+ * being matched against).
+ * @max_len
+ * The maximum permissible match length at this position. Must be >=
+ * BT_MATCHFINDER_REQUIRED_NBYTES.
* @nice_len
* Stop searching if a match of at least this length is found.
+ * Must be <= @max_len.
* @max_search_depth
- * Limit on the number of potential matches to consider.
- * @next_hash
- * Pointer to the hash code for the current sequence, which was computed
- * one position in advance so that the binary tree root could be
- * prefetched. This is an input/output parameter.
+ * Limit on the number of potential matches to consider. Must be >= 1.
+ * @next_hashes
+ * The precomputed hash codes for the sequence beginning at @in_next.
+ * These will be used and then updated with the precomputed hashcodes for
+ * the sequence beginning at @in_next + 1.
+ * @best_len_ret
+ * If a match of length >= 4 was found, then the length of the longest such
+ * match is written here; otherwise 3 is written here. (Note: this is
+ * redundant with the 'struct lz_match' array, but this is easier for the
+ * compiler to optimize when inlined and the caller immediately does a
+ * check against 'best_len'.)
+ * @lz_matchptr
+ * An array in which this function will record the matches. The recorded
+ * matches will be sorted by strictly increasing length and (non-strictly)
+ * increasing offset. The maximum number of matches that may be found is
+ * 'nice_len - 1', or one less if length 2 matches are disabled.
*
- * Note: this is very similar to bt_matchfinder_get_matches() because both
- * functions must do hashing and tree re-rooting. This version just doesn't
- * actually record any matches.
+ * The return value is a pointer to the next available slot in the @lz_matchptr
+ * array. (If no matches were found, this will be the same as @lz_matchptr.)
*/
-static inline void
-TEMPLATED(bt_matchfinder_skip_position)(struct TEMPLATED(bt_matchfinder) * const restrict mf,
- const u8 * const in_begin,
- const u8 * const in_next,
- const u8 * const in_end,
- const unsigned nice_len,
- const unsigned max_search_depth,
- u32 * restrict next_hash)
+static inline struct lz_match *
+TEMPLATED(bt_matchfinder_get_matches)(struct TEMPLATED(bt_matchfinder) *mf,
+ const u8 *in_begin,
+ ptrdiff_t cur_pos,
+ u32 max_len,
+ u32 nice_len,
+ u32 max_search_depth,
+ u32 next_hashes[static 2],
+ u32 *best_len_ret,
+ struct lz_match *lz_matchptr)
{
- unsigned depth_remaining = max_search_depth;
- u32 hash;
- pos_t cur_node;
- const u8 *matchptr;
- pos_t *pending_lt_ptr, *pending_gt_ptr;
- unsigned best_lt_len, best_gt_len;
- unsigned len;
-
- if (unlikely(in_end - in_next < LZ_HASH3_REQUIRED_NBYTES + 1))
- return;
-
- hash = *next_hash;
- *next_hash = bt_matchfinder_hash_3_bytes(in_next + 1);
- cur_node = mf->hash_tab[hash];
- mf->hash_tab[hash] = in_next - in_begin;
- prefetchw(&mf->hash_tab[*next_hash]);
-
- depth_remaining = max_search_depth;
- pending_lt_ptr = TEMPLATED(bt_left_child)(mf, in_next - in_begin);
- pending_gt_ptr = TEMPLATED(bt_right_child)(mf, in_next - in_begin);
- best_lt_len = 0;
- best_gt_len = 0;
- len = 0;
-
- if (!cur_node) {
- *pending_lt_ptr = 0;
- *pending_gt_ptr = 0;
- return;
- }
-
- for (;;) {
- matchptr = &in_begin[cur_node];
-
- if (matchptr[len] == in_next[len]) {
- len = lz_extend(in_next, matchptr, len + 1, nice_len);
- if (len == nice_len) {
- *pending_lt_ptr = *TEMPLATED(bt_left_child)(mf, cur_node);
- *pending_gt_ptr = *TEMPLATED(bt_right_child)(mf, cur_node);
- return;
- }
- }
-
- if (matchptr[len] < in_next[len]) {
- *pending_lt_ptr = cur_node;
- pending_lt_ptr = TEMPLATED(bt_right_child)(mf, cur_node);
- cur_node = *pending_lt_ptr;
- best_lt_len = len;
- if (best_gt_len < len)
- len = best_gt_len;
- } else {
- *pending_gt_ptr = cur_node;
- pending_gt_ptr = TEMPLATED(bt_left_child)(mf, cur_node);
- cur_node = *pending_gt_ptr;
- best_gt_len = len;
- if (best_lt_len < len)
- len = best_lt_len;
- }
+ return TEMPLATED(bt_matchfinder_advance_one_byte)(mf,
+ in_begin,
+ cur_pos,
+ max_len,
+ nice_len,
+ max_search_depth,
+ next_hashes,
+ best_len_ret,
+ lz_matchptr,
+ true);
+}
- if (!cur_node || !--depth_remaining) {
- *pending_lt_ptr = 0;
- *pending_gt_ptr = 0;
- return;
- }
- }
+/*
+ * Advance the matchfinder, but don't record any matches.
+ *
+ * This is very similar to bt_matchfinder_get_matches() because both functions
+ * must do hashing and tree re-rooting.
+ */
+static inline void
+TEMPLATED(bt_matchfinder_skip_position)(struct TEMPLATED(bt_matchfinder) *mf,
+ const u8 *in_begin,
+ ptrdiff_t cur_pos,
+ u32 max_len,
+ u32 nice_len,
+ u32 max_search_depth,
+ u32 next_hashes[static 2])
+{
+ u32 best_len;
+ TEMPLATED(bt_matchfinder_advance_one_byte)(mf,
+ in_begin,
+ cur_pos,
+ max_len,
+ nice_len,
+ max_search_depth,
+ next_hashes,
+ &best_len,
+ NULL,
+ false);
}
git://wimlib.net / wimlib / blobdiff