From: Robert Spalek Date: Mon, 14 Jun 2004 10:12:19 +0000 (+0000) Subject: sped up approximately 6 times: X-Git-Tag: holmes-import~1057 X-Git-Url: http://mj.ucw.cz/gitweb/?a=commitdiff_plain;h=e443a4773ae30fac352161cf3f9f6d97f229b8c3;p=libucw.git sped up approximately 6 times: - the whole idea of 2 hash-tables (for 3- and 4- matches) was bad - also, collision link-lists with errors were too bad ===> greatly simplified: only one hash-table/hash-function/link-list/... for 3-matches, double-linked link-list that can be maintained in constant time while preserving correctness, links to strings made implicit (hence the data structures is half-size and it fits better into the CPU-cache), no arithmetics when computing the hash-function, tuned constants determining the compression level, commented out code for 2-matches, ... --- diff --git a/lib/lizard.c b/lib/lizard.c index df215fdf..8d53234b 100644 --- a/lib/lizard.c +++ b/lib/lizard.c @@ -14,77 +14,53 @@ typedef u16 hash_ptr_t; struct hash_record { - byte *string; // in original text - hash_ptr_t next4, next3; // 0=end + /* the position in the original text is implicit; it is computed by locate_string() */ + hash_ptr_t next; // 0=end + hash_ptr_t prev; // high bit: 0=record in array, 1=head in hash-table (i.e. value of hashf) }; -#define HASH_SIZE ((1<<14) + 27) // prime, size of hash-table +#define HASH_SIZE (1<<14) // size of hash-table #define HASH_RECORDS (1<<15) // maximum number of records in hash-table, 0 is unused ==> subtract 1 -#define CHAIN_MAX_TESTS 16 // crop longer collision chains -#define CHAIN_GOOD_MATCH 16 // we already have a good match => end +#define CHAIN_MAX_TESTS 8 // crop longer collision chains +#define CHAIN_GOOD_MATCH 32 // we already have a good match => end static inline uns hashf(byte *string) + /* 0..HASH_SIZE-1 */ { - uns hash; -#ifdef CPU_ALLOW_UNALIGNED - hash = * (uns *) string; -#elif CPU_LITTLE_ENDIAN - hash = string[0] | (string[1]<<8) | (string[2]<<16) | (string[3]<<24); -#else - hash = string[3] | (string[2]<<8) | (string[1]<<16) | (string[0]<<24); -#endif - return hash; + return string[0] ^ (string[1]<<3) ^ (string[2]<<6); } -#ifdef CPU_LITTLE_ENDIAN -#define MASK_4TH_CHAR (~(0xff << 24)) -#else -#define MASK_4TH_CHAR (~0xff) -#endif -static inline uns -find_match(uns list_id, hash_ptr_t *hash_tab, struct hash_record *hash_rec, uns hash, byte *string, byte *string_end, byte **best_ptr) - /* - * We heavily rely on gcc optimisations (e.g. list_id is a constant hence - * some branches of the code will be pruned). - * - * hash_tab[hash] points to the head of the link-list of strings with - * the same hash. The records are statically stored in circular array - * hash_rec (with 1st entry unused), and the pointers are just 16-bit - * indices. In one array, we store 2 independent link-lists: with 3 and 4 - * hashed characters. - * - * The data structure fulfills 2 invariants: the strings in every collision - * chain are ordered by age, and the head of every link-list is placed - * correctly. The records in the middle of the link-lists may be placed - * wrong, and this is handled here by checking the string order. - */ +static inline byte * +locate_string(byte *string, uns record_id, uns head) + /* The strings are recorded into the hash-table regularly, hence there is no + * need to store the pointer there. */ { - if (list_id == 3) /* Find the 1st record in the collision chain. It is correct. */ - hash &= MASK_4TH_CHAR; - hash %= HASH_SIZE; - if (!hash_tab[hash]) - return 0; - struct hash_record *record = hash_rec + hash_tab[hash]; + string += record_id - head; + if (record_id >= head) + string -= HASH_RECORDS-1; + return string; +} +static inline uns +find_match(uns record_id, struct hash_record *hash_rec, byte *string, byte *string_end, byte **best_ptr, uns head) + /* hash_tab[hash] == record_id points to the head of the double-linked + * link-list of strings with the same hash. The records are statically + * stored in circular array hash_rec (with the 1st entry unused), and the + * pointers are just 16-bit indices. The strings in every collision chain + * are ordered by age. */ +{ uns count = CHAIN_MAX_TESTS; uns best_len = 0; - byte *last_cmp = string_end; - hash_ptr_t *last_ptr = NULL; - while (count-- > 0) + while (record_id && count-- > 0) { - byte *cmp = record->string; - if (cmp >= last_cmp) /* collision chain is ordered by age */ - { - *last_ptr = 0; /* prune */ - break; - } - last_cmp = cmp; + byte *record_string = locate_string(string, record_id, head); + byte *cmp = record_string; if (cmp[0] == string[0] && cmp[2] == string[2]) + /* implies cmp[1] == string[1] */ { if (cmp[3] == string[3]) { - /* implies cmp[1] == string[1], becase we hash into 14+eps bits */ cmp += 4; if (*cmp++ == string[4] && *cmp++ == string[5] && *cmp++ == string[6] && *cmp++ == string[7]) @@ -93,70 +69,46 @@ find_match(uns list_id, hash_ptr_t *hash_tab, struct hash_record *hash_rec, uns while (str <= string_end && *cmp++ == *str++); } } - else if (list_id == 3) - /* hashing 3 characters implies cmp[1] == string[1] - hashing 4 characters implies cmp[1] != string[1] */ - cmp += 4; else - goto next_collision; - uns len = cmp - record->string - 1; /* cmp points 2 characters after the last match */ + cmp += 4; + uns len = cmp - record_string - 1; /* cmp points 2 characters after the last match */ if (len > best_len) { best_len = len; - *best_ptr = record->string; - if (best_len >= CHAIN_GOOD_MATCH) /* optimisation */ + *best_ptr = record_string; + if (best_len >= CHAIN_GOOD_MATCH) /* optimization */ break; } } -next_collision: - if (list_id == 3) /* move to the next collision */ - last_ptr = &record->next3; - else - last_ptr = &record->next4; - uns next = *last_ptr; - if (!next) - break; - record = hash_rec + next; + record_id = hash_rec[record_id].next; } return best_len; } static uns -hash_string(byte *string, uns hash, hash_ptr_t *tab3, hash_ptr_t *tab4, struct hash_record *hash_rec, uns head) - /* - * Updates are performed more often than searches, so they are a bit lazy: we - * add new records in front of the link-lists, but only delete them when they - * are the _head_ of some link-list. This is done in constant time. - * - * By reusing an already allocated space in the circular array we necessarily - * alter records that are referred in the _middle_ of some link-list. - * However the newly inserted string is younger than anything else in the - * table and it can be checked easily. - */ +hash_string(hash_ptr_t *hash_tab, uns hash, struct hash_record *hash_rec, /*byte *string,*/ uns head, uns *to_delete) + /* We reuse hash-records stored in a circular array. First, delete the old + * one and then add the new one in front of the link-list. */ { struct hash_record *rec = hash_rec + head; - if (rec->string) /* unlink the original record if it was the _head_ of some link-list */ + if (*to_delete) /* unlink the original record */ { - uns h4 = hashf(rec->string); - uns h3 = h4 & MASK_4TH_CHAR; - h3 %= HASH_SIZE; - h4 %= HASH_SIZE; - if (tab3[h3] == head) - tab3[h3] = 0; - if (tab4[h4] == head) - tab4[h4] = 0; - + uns prev_id = rec->prev & ((1<<15)-1); + if (rec->prev & (1<<15)) /* was a head */ + hash_tab[prev_id] = 0; + else /* thanks to the ordering, this was a tail */ + hash_rec[prev_id].next = 0; } - uns h3 = (hash & MASK_4TH_CHAR) % HASH_SIZE; - uns h4 = hash % HASH_SIZE; - rec->string = string; /* is younger than any string inserted before */ - rec->next3 = tab3[h3]; - rec->next4 = tab4[h4]; - tab3[h3] = head; /* add the new record before the link-list */ - tab4[h4] = head; - head++; /* circular buffer, reuse old records, 0 is unused */ - if (head >= HASH_RECORDS) + rec->next = hash_tab[hash]; + rec->prev = (1<<15) | hash; + hash_rec[rec->next].prev = head; + hash_tab[hash] = head; /* add the new record before the link-list */ + + if (++head >= HASH_RECORDS) /* circular buffer, reuse old records, 0 is unused */ + { head = 1; + *to_delete = 1; + } return head; } @@ -202,65 +154,63 @@ lizzard_compress(byte *in, uns in_len, byte *out) * LIZZARD_MAX_ADD. There must be at least LIZZARD_NEEDS_CHARS characters * allocated after in. Returns the actual compressed length. */ { - hash_ptr_t hash_tab3[HASH_SIZE], hash_tab4[HASH_SIZE]; + hash_ptr_t hash_tab[HASH_SIZE]; struct hash_record hash_rec[HASH_RECORDS]; byte *in_start = in; byte *in_end = in + in_len; byte *out_start = out; byte *copy_start = in; uns head = 1; /* 0 in unused */ - bzero(hash_tab3, sizeof(hash_tab3)); /* init the hash-table */ - bzero(hash_tab4, sizeof(hash_tab4)); + uns to_delete = 0; + bzero(hash_tab, sizeof(hash_tab)); /* init the hash-table */ while (in < in_end) { uns hash = hashf(in); byte *best; - uns len = find_match(4, hash_tab4, hash_rec, hash, in, in_end, &best); - if (len >= 3) - goto match; - len = find_match(3, hash_tab3, hash_rec, hash, in, in_end, &best); - -match_found: - if (len >= 3) - goto match; + uns len = find_match(hash_tab[hash], hash_rec, in, in_end, &best, head); + if (len < 3) #if 0 // TODO: now, our routine does not detect matches of length 2 - if (len == 2 && (in - best->string) < ((1<<10) + 1)) - goto match; + if (len == 2 && (in - best->string) < ((1<<10) + 1)) + { /* pass-thru */ } + else #endif + { literal: - head = hash_string(in, hash, hash_tab3, hash_tab4, hash_rec, head); - in++; /* add a literal */ - continue; + head = hash_string(hash_tab, hash, hash_rec, head, &to_delete); + in++; /* add a literal */ + continue; + } -match: if (in + len > in_end) /* crop EOF */ { len = in_end - in; if (len < 3) - goto match_found; + goto literal; } /* Record the match. */ uns copy_len = in - copy_start; uns is_in_copy_mode = copy_start==in_start || copy_len >= 4; uns shift = in - best - 1; /* Try to use a 2-byte sequence. */ +#if 0 if (len == 2) { if (is_in_copy_mode || !copy_len) /* cannot use with 0 copied characters, because this bit pattern is reserved for copy mode */ goto literal; - + else + goto dump_2sequence; + } else +#endif + if (len == 3 && is_in_copy_mode && shift >= (1<<11) && shift < (1<<11) + (1<<10)) + { + /* optimisation for length-3 matches after a copy command */ + shift -= 1<<11; dump_2sequence: if (copy_len > 0) out = flush_copy_command(copy_start==in_start, out, copy_start, copy_len); *out++ = (shift>>6) & ~3; /* shift fits into 10 bits */ *out++ = shift & 0xff; } - else if (len == 3 && is_in_copy_mode && shift >= (1<<11) && shift < (1<<11) + (1<<10)) - { - /* optimisation for length-3 matches after a copy command */ - shift -= 1<<11; - goto dump_2sequence; - } /* now, len >= 3 */ else if (shift < (1<<11) && len <= 8) { @@ -301,9 +251,9 @@ dump_2sequence: *out++ = shift & 0xff; } /* Update the hash-table. */ - head = hash_string(in, hash, hash_tab3, hash_tab4, hash_rec, head); + head = hash_string(hash_tab, hash, hash_rec, head, &to_delete); for (uns i=1; i