1// Copyright 2006, 2008 Google Inc.
2// Authors: Chandra Chereddi, Lincoln Smith
3//
4// Licensed under the Apache License, Version 2.0 (the "License");
5// you may not use this file except in compliance with the License.
6// You may obtain a copy of the License at
7//
8//      http://www.apache.org/licenses/LICENSE-2.0
9//
10// Unless required by applicable law or agreed to in writing, software
11// distributed under the License is distributed on an "AS IS" BASIS,
12// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13// See the License for the specific language governing permissions and
14// limitations under the License.
15
16#include <config.h>
17#include "blockhash.h"
18#include "compile_assert.h"
19#include <stdint.h>  // uint32_t
20#include <string.h>  // memcpy, memcmp
21#include "logging.h"
22#include "rolling_hash.h"
23
24namespace open_vcdiff {
25
26typedef unsigned long uword_t;  // a machine word                         NOLINT
27
28BlockHash::BlockHash(const char* source_data,
29                     size_t source_size,
30                     int starting_offset)
31    : source_data_(source_data),
32      source_size_(source_size),
33      hash_table_mask_(0),
34      starting_offset_(starting_offset),
35      last_block_added_(-1) {
36}
37
38BlockHash::~BlockHash() { }
39
40// kBlockSize must be at least 2 to be meaningful.  Since it's a compile-time
41// constant, check its value at compile time rather than wasting CPU cycles
42// on runtime checks.
43COMPILE_ASSERT(BlockHash::kBlockSize >= 2, kBlockSize_must_be_at_least_2);
44
45// kBlockSize is required to be a power of 2 because multiplication
46// (n * kBlockSize), division (n / kBlockSize) and MOD (n % kBlockSize)
47// are commonly-used operations.  If kBlockSize is a compile-time
48// constant and a power of 2, the compiler can convert these three operations
49// into bit-shift (>> or <<) and bitwise-AND (&) operations, which are much
50// more efficient than executing full integer multiply, divide, or remainder
51// instructions.
52COMPILE_ASSERT((BlockHash::kBlockSize & (BlockHash::kBlockSize - 1)) == 0,
53               kBlockSize_must_be_a_power_of_2);
54
55bool BlockHash::Init(bool populate_hash_table) {
56  if (!hash_table_.empty() ||
57      !next_block_table_.empty() ||
58      !last_block_table_.empty()) {
59    LOG(DFATAL) << "Init() called twice for same BlockHash object" << LOG_ENDL;
60    return false;
61  }
62  const size_t table_size = CalcTableSize(source_size_);
63  if (table_size == 0) {
64    LOG(DFATAL) << "Error finding table size for source size " << source_size_
65                << LOG_ENDL;
66    return false;
67  }
68  // Since table_size is a power of 2, (table_size - 1) is a bit mask
69  // containing all the bits below table_size.
70  hash_table_mask_ = static_cast<uint32_t>(table_size - 1);
71  hash_table_.resize(table_size, -1);
72  next_block_table_.resize(GetNumberOfBlocks(), -1);
73  last_block_table_.resize(GetNumberOfBlocks(), -1);
74  if (populate_hash_table) {
75    AddAllBlocks();
76  }
77  return true;
78}
79
80const BlockHash* BlockHash::CreateDictionaryHash(const char* dictionary_data,
81                                                 size_t dictionary_size) {
82  BlockHash* new_dictionary_hash = new BlockHash(dictionary_data,
83                                                 dictionary_size,
84                                                 0);
85  if (!new_dictionary_hash->Init(/* populate_hash_table = */ true)) {
86    delete new_dictionary_hash;
87    return NULL;
88  } else {
89    return new_dictionary_hash;
90  }
91}
92
93BlockHash* BlockHash::CreateTargetHash(const char* target_data,
94                                       size_t target_size,
95                                       size_t dictionary_size) {
96  BlockHash* new_target_hash = new BlockHash(target_data,
97                                             target_size,
98                                             static_cast<int>(dictionary_size));
99  if (!new_target_hash->Init(/* populate_hash_table = */ false)) {
100    delete new_target_hash;
101    return NULL;
102  } else {
103    return new_target_hash;
104  }
105}
106
107// Returns zero if an error occurs.
108size_t BlockHash::CalcTableSize(const size_t dictionary_size) {
109  // Overallocate the hash table by making it the same size (in bytes)
110  // as the source data.  This is a trade-off between space and time:
111  // the empty entries in the hash table will reduce the
112  // probability of a hash collision to (sizeof(int) / kblockSize),
113  // and so save time comparing false matches.
114  const size_t min_size = (dictionary_size / sizeof(int)) + 1;  // NOLINT
115  size_t table_size = 1;
116  // Find the smallest power of 2 that is >= min_size, and assign
117  // that value to table_size.
118  while (table_size < min_size) {
119    table_size <<= 1;
120    // Guard against an infinite loop
121    if (table_size <= 0) {
122      LOG(DFATAL) << "Internal error: CalcTableSize(dictionary_size = "
123                  << dictionary_size
124                  << "): resulting table_size " << table_size
125                  << " is zero or negative" << LOG_ENDL;
126      return 0;
127    }
128  }
129  // Check size sanity
130  if ((table_size & (table_size - 1)) != 0) {
131    LOG(DFATAL) << "Internal error: CalcTableSize(dictionary_size = "
132                << dictionary_size
133                << "): resulting table_size " << table_size
134                << " is not a power of 2" << LOG_ENDL;
135    return 0;
136  }
137  // The loop above tries to find the smallest power of 2 that is >= min_size.
138  // That value must lie somewhere between min_size and (min_size * 2),
139  // except for the case (dictionary_size == 0, table_size == 1).
140  if ((dictionary_size > 0) && (table_size > (min_size * 2))) {
141    LOG(DFATAL) << "Internal error: CalcTableSize(dictionary_size = "
142                << dictionary_size
143                << "): resulting table_size " << table_size
144                << " is too large" << LOG_ENDL;
145    return 0;
146  }
147  return table_size;
148}
149
150// If the hash value is already available from the rolling hash,
151// call this function to save time.
152void BlockHash::AddBlock(uint32_t hash_value) {
153  if (hash_table_.empty()) {
154    LOG(DFATAL) << "BlockHash::AddBlock() called before BlockHash::Init()"
155                << LOG_ENDL;
156    return;
157  }
158  // The initial value of last_block_added_ is -1.
159  int block_number = last_block_added_ + 1;
160  const int total_blocks =
161      static_cast<int>(source_size_ / kBlockSize);  // round down
162  if (block_number >= total_blocks) {
163    LOG(DFATAL) << "BlockHash::AddBlock() called"
164                   " with block number " << block_number
165                << " that is past last block " << (total_blocks - 1)
166                << LOG_ENDL;
167    return;
168  }
169  if (next_block_table_[block_number] != -1) {
170    LOG(DFATAL) << "Internal error in BlockHash::AddBlock(): "
171                   "block number = " << block_number
172                << ", next block should be -1 but is "
173                << next_block_table_[block_number] << LOG_ENDL;
174    return;
175  }
176  const uint32_t hash_table_index = GetHashTableIndex(hash_value);
177  const int first_matching_block = hash_table_[hash_table_index];
178  if (first_matching_block < 0) {
179    // This is the first entry with this hash value
180    hash_table_[hash_table_index] = block_number;
181    last_block_table_[block_number] = block_number;
182  } else {
183    // Add this entry at the end of the chain of matching blocks
184    const int last_matching_block = last_block_table_[first_matching_block];
185    if (next_block_table_[last_matching_block] != -1) {
186      LOG(DFATAL) << "Internal error in BlockHash::AddBlock(): "
187                     "first matching block = " << first_matching_block
188                  << ", last matching block = " << last_matching_block
189                  << ", next block should be -1 but is "
190                  << next_block_table_[last_matching_block] << LOG_ENDL;
191      return;
192    }
193    next_block_table_[last_matching_block] = block_number;
194    last_block_table_[first_matching_block] = block_number;
195  }
196  last_block_added_ = block_number;
197}
198
199void BlockHash::AddAllBlocks() {
200  AddAllBlocksThroughIndex(static_cast<int>(source_size_));
201}
202
203void BlockHash::AddAllBlocksThroughIndex(int end_index) {
204  if (end_index > static_cast<int>(source_size_)) {
205    LOG(DFATAL) << "BlockHash::AddAllBlocksThroughIndex() called"
206                   " with index " << end_index
207                << " higher than end index  " << source_size_ << LOG_ENDL;
208    return;
209  }
210  const int last_index_added = last_block_added_ * kBlockSize;
211  if (end_index <= last_index_added) {
212    LOG(DFATAL) << "BlockHash::AddAllBlocksThroughIndex() called"
213                   " with index " << end_index
214                << " <= last index added ( " << last_index_added
215                << ")" << LOG_ENDL;
216    return;
217  }
218  int end_limit = end_index;
219  // Don't allow reading any indices at or past source_size_.
220  // The Hash function extends (kBlockSize - 1) bytes past the index,
221  // so leave a margin of that size.
222  int last_legal_hash_index = static_cast<int>(source_size() - kBlockSize);
223  if (end_limit > last_legal_hash_index) {
224    end_limit = last_legal_hash_index + 1;
225  }
226  const char* block_ptr = source_data() + NextIndexToAdd();
227  const char* const end_ptr = source_data() + end_limit;
228  while (block_ptr < end_ptr) {
229    AddBlock(RollingHash<kBlockSize>::Hash(block_ptr));
230    block_ptr += kBlockSize;
231  }
232}
233
234COMPILE_ASSERT((BlockHash::kBlockSize % sizeof(uword_t)) == 0,
235               kBlockSize_must_be_a_multiple_of_machine_word_size);
236
237// A recursive template to compare a fixed number
238// of (possibly unaligned) machine words starting
239// at addresses block1 and block2.  Returns true or false
240// depending on whether an exact match was found.
241template<int number_of_words>
242inline bool CompareWholeWordValues(const char* block1,
243                                   const char* block2) {
244  return CompareWholeWordValues<1>(block1, block2) &&
245         CompareWholeWordValues<number_of_words - 1>(block1 + sizeof(uword_t),
246                                                     block2 + sizeof(uword_t));
247}
248
249// The base of the recursive template: compare one pair of machine words.
250template<>
251inline bool CompareWholeWordValues<1>(const char* word1,
252                                      const char* word2) {
253  uword_t aligned_word1, aligned_word2;
254  memcpy(&aligned_word1, word1, sizeof(aligned_word1));
255  memcpy(&aligned_word2, word2, sizeof(aligned_word2));
256  return aligned_word1 == aligned_word2;
257}
258
259// A block must be composed of an integral number of machine words
260// (uword_t values.)  This function takes advantage of that fact
261// by comparing the blocks as series of (possibly unaligned) word values.
262// A word-sized comparison can be performed as a single
263// machine instruction.  Comparing words instead of bytes means that,
264// on a 64-bit platform, this function will use 8 times fewer test-and-branch
265// instructions than a byte-by-byte comparison.  Even with the extra
266// cost of the calls to memcpy, this method is still at least twice as fast
267// as memcmp (measured using gcc on a 64-bit platform, with a block size
268// of 32.)  For blocks with identical contents (a common case), this method
269// is over six times faster than memcmp.
270inline bool BlockCompareWordsInline(const char* block1, const char* block2) {
271  static const size_t kWordsPerBlock = BlockHash::kBlockSize / sizeof(uword_t);
272  return CompareWholeWordValues<kWordsPerBlock>(block1, block2);
273}
274
275bool BlockHash::BlockCompareWords(const char* block1, const char* block2) {
276  return BlockCompareWordsInline(block1, block2);
277}
278
279inline bool BlockContentsMatchInline(const char* block1, const char* block2) {
280  // Optimize for mismatch in first byte.  Since this function is called only
281  // when the hash values of the two blocks match, it is very likely that either
282  // the blocks are identical, or else the first byte does not match.
283  if (*block1 != *block2) {
284    return false;
285  }
286#ifdef VCDIFF_USE_BLOCK_COMPARE_WORDS
287  return BlockCompareWordsInline(block1, block2);
288#else  // !VCDIFF_USE_BLOCK_COMPARE_WORDS
289  return memcmp(block1, block2, BlockHash::kBlockSize) == 0;
290#endif  // VCDIFF_USE_BLOCK_COMPARE_WORDS
291}
292
293bool BlockHash::BlockContentsMatch(const char* block1, const char* block2) {
294  return BlockContentsMatchInline(block1, block2);
295}
296
297inline int BlockHash::SkipNonMatchingBlocks(int block_number,
298                                            const char* block_ptr) const {
299  int probes = 0;
300  while ((block_number >= 0) &&
301         !BlockContentsMatchInline(block_ptr,
302                                   &source_data_[block_number * kBlockSize])) {
303    if (++probes > kMaxProbes) {
304      return -1;  // Avoid too much chaining
305    }
306    block_number = next_block_table_[block_number];
307  }
308  return block_number;
309}
310
311// Init() must have been called and returned true before using
312// FirstMatchingBlock or NextMatchingBlock.  No check is performed
313// for this condition; the code will crash if this condition is violated.
314inline int BlockHash::FirstMatchingBlockInline(uint32_t hash_value,
315                                               const char* block_ptr) const {
316  return SkipNonMatchingBlocks(hash_table_[GetHashTableIndex(hash_value)],
317                               block_ptr);
318}
319
320int BlockHash::FirstMatchingBlock(uint32_t hash_value,
321                                  const char* block_ptr) const {
322  return FirstMatchingBlockInline(hash_value, block_ptr);
323}
324
325int BlockHash::NextMatchingBlock(int block_number,
326                                 const char* block_ptr) const {
327  if (static_cast<size_t>(block_number) >= GetNumberOfBlocks()) {
328    LOG(DFATAL) << "NextMatchingBlock called for invalid block number "
329                << block_number << LOG_ENDL;
330    return -1;
331  }
332  return SkipNonMatchingBlocks(next_block_table_[block_number], block_ptr);
333}
334
335// Keep a count of the number of matches found.  This will throttle the
336// number of iterations in FindBestMatch.  For example, if the entire
337// dictionary is made up of spaces (' ') and the search string is also
338// made up of spaces, there will be one match for each block in the
339// dictionary.
340inline bool BlockHash::TooManyMatches(int* match_counter) {
341  ++(*match_counter);
342  return (*match_counter) > kMaxMatchesToCheck;
343}
344
345// Returns the number of bytes to the left of source_match_start
346// that match the corresponding bytes to the left of target_match_start.
347// Will not examine more than max_bytes bytes, which is to say that
348// the return value will be in the range [0, max_bytes] inclusive.
349int BlockHash::MatchingBytesToLeft(const char* source_match_start,
350                                   const char* target_match_start,
351                                   int max_bytes) {
352  const char* source_ptr = source_match_start;
353  const char* target_ptr = target_match_start;
354  int bytes_found = 0;
355  while (bytes_found < max_bytes) {
356    --source_ptr;
357    --target_ptr;
358    if (*source_ptr != *target_ptr) {
359      break;
360    }
361    ++bytes_found;
362  }
363  return bytes_found;
364}
365
366// Returns the number of bytes starting at source_match_end
367// that match the corresponding bytes starting at target_match_end.
368// Will not examine more than max_bytes bytes, which is to say that
369// the return value will be in the range [0, max_bytes] inclusive.
370int BlockHash::MatchingBytesToRight(const char* source_match_end,
371                                    const char* target_match_end,
372                                    int max_bytes) {
373  const char* source_ptr = source_match_end;
374  const char* target_ptr = target_match_end;
375  int bytes_found = 0;
376  while ((bytes_found < max_bytes) && (*source_ptr == *target_ptr)) {
377    ++bytes_found;
378    ++source_ptr;
379    ++target_ptr;
380  }
381  return bytes_found;
382}
383
384// No NULL checks are performed on the pointer arguments.  The caller
385// must guarantee that none of the arguments is NULL, or a crash will occur.
386//
387// The vast majority of calls to FindBestMatch enter the loop *zero* times,
388// which is to say that most candidate blocks find no matches in the dictionary.
389// The important sections for optimization are therefore the code outside the
390// loop and the code within the loop conditions.  Keep this to a minimum.
391void BlockHash::FindBestMatch(uint32_t hash_value,
392                              const char* target_candidate_start,
393                              const char* target_start,
394                              size_t target_size,
395                              Match* best_match) const {
396  int match_counter = 0;
397  for (int block_number = FirstMatchingBlockInline(hash_value,
398                                                   target_candidate_start);
399       (block_number >= 0) && !TooManyMatches(&match_counter);
400       block_number = NextMatchingBlock(block_number, target_candidate_start)) {
401    int source_match_offset = block_number * kBlockSize;
402    const int source_match_end = source_match_offset + kBlockSize;
403
404    int target_match_offset =
405        static_cast<int>(target_candidate_start - target_start);
406    const int target_match_end = target_match_offset + kBlockSize;
407
408    size_t match_size = kBlockSize;
409    {
410      // Extend match start towards beginning of unencoded data
411      const int limit_bytes_to_left = std::min(source_match_offset,
412                                               target_match_offset);
413      const int matching_bytes_to_left =
414          MatchingBytesToLeft(source_data_ + source_match_offset,
415                              target_start + target_match_offset,
416                              limit_bytes_to_left);
417      source_match_offset -= matching_bytes_to_left;
418      target_match_offset -= matching_bytes_to_left;
419      match_size += matching_bytes_to_left;
420    }
421    {
422      // Extend match end towards end of unencoded data
423      const size_t source_bytes_to_right = source_size_ - source_match_end;
424      const size_t target_bytes_to_right = target_size - target_match_end;
425      const size_t limit_bytes_to_right = std::min(source_bytes_to_right,
426                                                   target_bytes_to_right);
427      match_size +=
428          MatchingBytesToRight(source_data_ + source_match_end,
429                               target_start + target_match_end,
430                               static_cast<int>(limit_bytes_to_right));
431    }
432    // Update in/out parameter if the best match found was better
433    // than any match already stored in *best_match.
434    best_match->ReplaceIfBetterMatch(match_size,
435                                     source_match_offset + starting_offset_,
436                                     target_match_offset);
437  }
438}
439
440}  // namespace open_vcdiff
441