Merge branch 'ds/commit-graph-incremental'
[gitweb.git] / blame.c
diff --git a/blame.c b/blame.c
index 145eaf2faf9cf56977da61572c93783ea702b0f9..7f04580ad57a67be099fb1c0d85340cbcdb1f70c 100644 (file)
--- a/blame.c
+++ b/blame.c
@@ -311,12 +311,707 @@ static int diff_hunks(mmfile_t *file_a, mmfile_t *file_b,
        return xdi_diff(file_a, file_b, &xpp, &xecfg, &ecb);
 }
 
+static const char *get_next_line(const char *start, const char *end)
+{
+       const char *nl = memchr(start, '\n', end - start);
+
+       return nl ? nl + 1 : end;
+}
+
+static int find_line_starts(int **line_starts, const char *buf,
+                           unsigned long len)
+{
+       const char *end = buf + len;
+       const char *p;
+       int *lineno;
+       int num = 0;
+
+       for (p = buf; p < end; p = get_next_line(p, end))
+               num++;
+
+       ALLOC_ARRAY(*line_starts, num + 1);
+       lineno = *line_starts;
+
+       for (p = buf; p < end; p = get_next_line(p, end))
+               *lineno++ = p - buf;
+
+       *lineno = len;
+
+       return num;
+}
+
+struct fingerprint_entry;
+
+/* A fingerprint is intended to loosely represent a string, such that two
+ * fingerprints can be quickly compared to give an indication of the similarity
+ * of the strings that they represent.
+ *
+ * A fingerprint is represented as a multiset of the lower-cased byte pairs in
+ * the string that it represents. Whitespace is added at each end of the
+ * string. Whitespace pairs are ignored. Whitespace is converted to '\0'.
+ * For example, the string "Darth   Radar" will be converted to the following
+ * fingerprint:
+ * {"\0d", "da", "da", "ar", "ar", "rt", "th", "h\0", "\0r", "ra", "ad", "r\0"}
+ *
+ * The similarity between two fingerprints is the size of the intersection of
+ * their multisets, including repeated elements. See fingerprint_similarity for
+ * examples.
+ *
+ * For ease of implementation, the fingerprint is implemented as a map
+ * of byte pairs to the count of that byte pair in the string, instead of
+ * allowing repeated elements in a set.
+ */
+struct fingerprint {
+       struct hashmap map;
+       /* As we know the maximum number of entries in advance, it's
+        * convenient to store the entries in a single array instead of having
+        * the hashmap manage the memory.
+        */
+       struct fingerprint_entry *entries;
+};
+
+/* A byte pair in a fingerprint. Stores the number of times the byte pair
+ * occurs in the string that the fingerprint represents.
+ */
+struct fingerprint_entry {
+       /* The hashmap entry - the hash represents the byte pair in its
+        * entirety so we don't need to store the byte pair separately.
+        */
+       struct hashmap_entry entry;
+       /* The number of times the byte pair occurs in the string that the
+        * fingerprint represents.
+        */
+       int count;
+};
+
+/* See `struct fingerprint` for an explanation of what a fingerprint is.
+ * \param result the fingerprint of the string is stored here. This must be
+ *              freed later using free_fingerprint.
+ * \param line_begin the start of the string
+ * \param line_end the end of the string
+ */
+static void get_fingerprint(struct fingerprint *result,
+                           const char *line_begin,
+                           const char *line_end)
+{
+       unsigned int hash, c0 = 0, c1;
+       const char *p;
+       int max_map_entry_count = 1 + line_end - line_begin;
+       struct fingerprint_entry *entry = xcalloc(max_map_entry_count,
+               sizeof(struct fingerprint_entry));
+       struct fingerprint_entry *found_entry;
+
+       hashmap_init(&result->map, NULL, NULL, max_map_entry_count);
+       result->entries = entry;
+       for (p = line_begin; p <= line_end; ++p, c0 = c1) {
+               /* Always terminate the string with whitespace.
+                * Normalise whitespace to 0, and normalise letters to
+                * lower case. This won't work for multibyte characters but at
+                * worst will match some unrelated characters.
+                */
+               if ((p == line_end) || isspace(*p))
+                       c1 = 0;
+               else
+                       c1 = tolower(*p);
+               hash = c0 | (c1 << 8);
+               /* Ignore whitespace pairs */
+               if (hash == 0)
+                       continue;
+               hashmap_entry_init(entry, hash);
+
+               found_entry = hashmap_get(&result->map, entry, NULL);
+               if (found_entry) {
+                       found_entry->count += 1;
+               } else {
+                       entry->count = 1;
+                       hashmap_add(&result->map, entry);
+                       ++entry;
+               }
+       }
+}
+
+static void free_fingerprint(struct fingerprint *f)
+{
+       hashmap_free(&f->map, 0);
+       free(f->entries);
+}
+
+/* Calculates the similarity between two fingerprints as the size of the
+ * intersection of their multisets, including repeated elements. See
+ * `struct fingerprint` for an explanation of the fingerprint representation.
+ * The similarity between "cat mat" and "father rather" is 2 because "at" is
+ * present twice in both strings while the similarity between "tim" and "mit"
+ * is 0.
+ */
+static int fingerprint_similarity(struct fingerprint *a, struct fingerprint *b)
+{
+       int intersection = 0;
+       struct hashmap_iter iter;
+       const struct fingerprint_entry *entry_a, *entry_b;
+
+       hashmap_iter_init(&b->map, &iter);
+
+       while ((entry_b = hashmap_iter_next(&iter))) {
+               if ((entry_a = hashmap_get(&a->map, entry_b, NULL))) {
+                       intersection += entry_a->count < entry_b->count ?
+                                       entry_a->count : entry_b->count;
+               }
+       }
+       return intersection;
+}
+
+/* Subtracts byte-pair elements in B from A, modifying A in place.
+ */
+static void fingerprint_subtract(struct fingerprint *a, struct fingerprint *b)
+{
+       struct hashmap_iter iter;
+       struct fingerprint_entry *entry_a;
+       const struct fingerprint_entry *entry_b;
+
+       hashmap_iter_init(&b->map, &iter);
+
+       while ((entry_b = hashmap_iter_next(&iter))) {
+               if ((entry_a = hashmap_get(&a->map, entry_b, NULL))) {
+                       if (entry_a->count <= entry_b->count)
+                               hashmap_remove(&a->map, entry_b, NULL);
+                       else
+                               entry_a->count -= entry_b->count;
+               }
+       }
+}
+
+/* Calculate fingerprints for a series of lines.
+ * Puts the fingerprints in the fingerprints array, which must have been
+ * preallocated to allow storing line_count elements.
+ */
+static void get_line_fingerprints(struct fingerprint *fingerprints,
+                                 const char *content, const int *line_starts,
+                                 long first_line, long line_count)
+{
+       int i;
+       const char *linestart, *lineend;
+
+       line_starts += first_line;
+       for (i = 0; i < line_count; ++i) {
+               linestart = content + line_starts[i];
+               lineend = content + line_starts[i + 1];
+               get_fingerprint(fingerprints + i, linestart, lineend);
+       }
+}
+
+static void free_line_fingerprints(struct fingerprint *fingerprints,
+                                  int nr_fingerprints)
+{
+       int i;
+
+       for (i = 0; i < nr_fingerprints; i++)
+               free_fingerprint(&fingerprints[i]);
+}
+
+/* This contains the data necessary to linearly map a line number in one half
+ * of a diff chunk to the line in the other half of the diff chunk that is
+ * closest in terms of its position as a fraction of the length of the chunk.
+ */
+struct line_number_mapping {
+       int destination_start, destination_length,
+               source_start, source_length;
+};
+
+/* Given a line number in one range, offset and scale it to map it onto the
+ * other range.
+ * Essentially this mapping is a simple linear equation but the calculation is
+ * more complicated to allow performing it with integer operations.
+ * Another complication is that if a line could map onto many lines in the
+ * destination range then we want to choose the line at the center of those
+ * possibilities.
+ * Example: if the chunk is 2 lines long in A and 10 lines long in B then the
+ * first 5 lines in B will map onto the first line in the A chunk, while the
+ * last 5 lines will all map onto the second line in the A chunk.
+ * Example: if the chunk is 10 lines long in A and 2 lines long in B then line
+ * 0 in B will map onto line 2 in A, and line 1 in B will map onto line 7 in A.
+ */
+static int map_line_number(int line_number,
+       const struct line_number_mapping *mapping)
+{
+       return ((line_number - mapping->source_start) * 2 + 1) *
+              mapping->destination_length /
+              (mapping->source_length * 2) +
+              mapping->destination_start;
+}
+
+/* Get a pointer to the element storing the similarity between a line in A
+ * and a line in B.
+ *
+ * The similarities are stored in a 2-dimensional array. Each "row" in the
+ * array contains the similarities for a line in B. The similarities stored in
+ * a row are the similarities between the line in B and the nearby lines in A.
+ * To keep the length of each row the same, it is padded out with values of -1
+ * where the search range extends beyond the lines in A.
+ * For example, if max_search_distance_a is 2 and the two sides of a diff chunk
+ * look like this:
+ * a | m
+ * b | n
+ * c | o
+ * d | p
+ * e | q
+ * Then the similarity array will contain:
+ * [-1, -1, am, bm, cm,
+ *  -1, an, bn, cn, dn,
+ *  ao, bo, co, do, eo,
+ *  bp, cp, dp, ep, -1,
+ *  cq, dq, eq, -1, -1]
+ * Where similarities are denoted either by -1 for invalid, or the
+ * concatenation of the two lines in the diff being compared.
+ *
+ * \param similarities array of similarities between lines in A and B
+ * \param line_a the index of the line in A, in the same frame of reference as
+ *     closest_line_a.
+ * \param local_line_b the index of the line in B, relative to the first line
+ *                    in B that similarities represents.
+ * \param closest_line_a the index of the line in A that is deemed to be
+ *                      closest to local_line_b. This must be in the same
+ *                      frame of reference as line_a. This value defines
+ *                      where similarities is centered for the line in B.
+ * \param max_search_distance_a maximum distance in lines from the closest line
+ *                             in A for other lines in A for which
+ *                             similarities may be calculated.
+ */
+static int *get_similarity(int *similarities,
+                          int line_a, int local_line_b,
+                          int closest_line_a, int max_search_distance_a)
+{
+       assert(abs(line_a - closest_line_a) <=
+              max_search_distance_a);
+       return similarities + line_a - closest_line_a +
+              max_search_distance_a +
+              local_line_b * (max_search_distance_a * 2 + 1);
+}
+
+#define CERTAIN_NOTHING_MATCHES -2
+#define CERTAINTY_NOT_CALCULATED -1
+
+/* Given a line in B, first calculate its similarities with nearby lines in A
+ * if not already calculated, then identify the most similar and second most
+ * similar lines. The "certainty" is calculated based on those two
+ * similarities.
+ *
+ * \param start_a the index of the first line of the chunk in A
+ * \param length_a the length in lines of the chunk in A
+ * \param local_line_b the index of the line in B, relative to the first line
+ *                    in the chunk.
+ * \param fingerprints_a array of fingerprints for the chunk in A
+ * \param fingerprints_b array of fingerprints for the chunk in B
+ * \param similarities 2-dimensional array of similarities between lines in A
+ *                    and B. See get_similarity() for more details.
+ * \param certainties array of values indicating how strongly a line in B is
+ *                   matched with some line in A.
+ * \param second_best_result array of absolute indices in A for the second
+ *                          closest match of a line in B.
+ * \param result array of absolute indices in A for the closest match of a line
+ *              in B.
+ * \param max_search_distance_a maximum distance in lines from the closest line
+ *                             in A for other lines in A for which
+ *                             similarities may be calculated.
+ * \param map_line_number_in_b_to_a parameter to map_line_number().
+ */
+static void find_best_line_matches(
+       int start_a,
+       int length_a,
+       int start_b,
+       int local_line_b,
+       struct fingerprint *fingerprints_a,
+       struct fingerprint *fingerprints_b,
+       int *similarities,
+       int *certainties,
+       int *second_best_result,
+       int *result,
+       const int max_search_distance_a,
+       const struct line_number_mapping *map_line_number_in_b_to_a)
+{
+
+       int i, search_start, search_end, closest_local_line_a, *similarity,
+               best_similarity = 0, second_best_similarity = 0,
+               best_similarity_index = 0, second_best_similarity_index = 0;
+
+       /* certainty has already been calculated so no need to redo the work */
+       if (certainties[local_line_b] != CERTAINTY_NOT_CALCULATED)
+               return;
+
+       closest_local_line_a = map_line_number(
+               local_line_b + start_b, map_line_number_in_b_to_a) - start_a;
+
+       search_start = closest_local_line_a - max_search_distance_a;
+       if (search_start < 0)
+               search_start = 0;
+
+       search_end = closest_local_line_a + max_search_distance_a + 1;
+       if (search_end > length_a)
+               search_end = length_a;
+
+       for (i = search_start; i < search_end; ++i) {
+               similarity = get_similarity(similarities,
+                                           i, local_line_b,
+                                           closest_local_line_a,
+                                           max_search_distance_a);
+               if (*similarity == -1) {
+                       /* This value will never exceed 10 but assert just in
+                        * case
+                        */
+                       assert(abs(i - closest_local_line_a) < 1000);
+                       /* scale the similarity by (1000 - distance from
+                        * closest line) to act as a tie break between lines
+                        * that otherwise are equally similar.
+                        */
+                       *similarity = fingerprint_similarity(
+                               fingerprints_b + local_line_b,
+                               fingerprints_a + i) *
+                               (1000 - abs(i - closest_local_line_a));
+               }
+               if (*similarity > best_similarity) {
+                       second_best_similarity = best_similarity;
+                       second_best_similarity_index = best_similarity_index;
+                       best_similarity = *similarity;
+                       best_similarity_index = i;
+               } else if (*similarity > second_best_similarity) {
+                       second_best_similarity = *similarity;
+                       second_best_similarity_index = i;
+               }
+       }
+
+       if (best_similarity == 0) {
+               /* this line definitely doesn't match with anything. Mark it
+                * with this special value so it doesn't get invalidated and
+                * won't be recalculated.
+                */
+               certainties[local_line_b] = CERTAIN_NOTHING_MATCHES;
+               result[local_line_b] = -1;
+       } else {
+               /* Calculate the certainty with which this line matches.
+                * If the line matches well with two lines then that reduces
+                * the certainty. However we still want to prioritise matching
+                * a line that matches very well with two lines over matching a
+                * line that matches poorly with one line, hence doubling
+                * best_similarity.
+                * This means that if we have
+                * line X that matches only one line with a score of 3,
+                * line Y that matches two lines equally with a score of 5,
+                * and line Z that matches only one line with a score or 2,
+                * then the lines in order of certainty are X, Y, Z.
+                */
+               certainties[local_line_b] = best_similarity * 2 -
+                       second_best_similarity;
+
+               /* We keep both the best and second best results to allow us to
+                * check at a later stage of the matching process whether the
+                * result needs to be invalidated.
+                */
+               result[local_line_b] = start_a + best_similarity_index;
+               second_best_result[local_line_b] =
+                       start_a + second_best_similarity_index;
+       }
+}
+
+/*
+ * This finds the line that we can match with the most confidence, and
+ * uses it as a partition. It then calls itself on the lines on either side of
+ * that partition. In this way we avoid lines appearing out of order, and
+ * retain a sensible line ordering.
+ * \param start_a index of the first line in A with which lines in B may be
+ *               compared.
+ * \param start_b index of the first line in B for which matching should be
+ *               done.
+ * \param length_a number of lines in A with which lines in B may be compared.
+ * \param length_b number of lines in B for which matching should be done.
+ * \param fingerprints_a mutable array of fingerprints in A. The first element
+ *                      corresponds to the line at start_a.
+ * \param fingerprints_b array of fingerprints in B. The first element
+ *                      corresponds to the line at start_b.
+ * \param similarities 2-dimensional array of similarities between lines in A
+ *                    and B. See get_similarity() for more details.
+ * \param certainties array of values indicating how strongly a line in B is
+ *                   matched with some line in A.
+ * \param second_best_result array of absolute indices in A for the second
+ *                          closest match of a line in B.
+ * \param result array of absolute indices in A for the closest match of a line
+ *              in B.
+ * \param max_search_distance_a maximum distance in lines from the closest line
+ *                           in A for other lines in A for which
+ *                           similarities may be calculated.
+ * \param max_search_distance_b an upper bound on the greatest possible
+ *                           distance between lines in B such that they will
+ *                              both be compared with the same line in A
+ *                           according to max_search_distance_a.
+ * \param map_line_number_in_b_to_a parameter to map_line_number().
+ */
+static void fuzzy_find_matching_lines_recurse(
+       int start_a, int start_b,
+       int length_a, int length_b,
+       struct fingerprint *fingerprints_a,
+       struct fingerprint *fingerprints_b,
+       int *similarities,
+       int *certainties,
+       int *second_best_result,
+       int *result,
+       int max_search_distance_a,
+       int max_search_distance_b,
+       const struct line_number_mapping *map_line_number_in_b_to_a)
+{
+       int i, invalidate_min, invalidate_max, offset_b,
+               second_half_start_a, second_half_start_b,
+               second_half_length_a, second_half_length_b,
+               most_certain_line_a, most_certain_local_line_b = -1,
+               most_certain_line_certainty = -1,
+               closest_local_line_a;
+
+       for (i = 0; i < length_b; ++i) {
+               find_best_line_matches(start_a,
+                                      length_a,
+                                      start_b,
+                                      i,
+                                      fingerprints_a,
+                                      fingerprints_b,
+                                      similarities,
+                                      certainties,
+                                      second_best_result,
+                                      result,
+                                      max_search_distance_a,
+                                      map_line_number_in_b_to_a);
+
+               if (certainties[i] > most_certain_line_certainty) {
+                       most_certain_line_certainty = certainties[i];
+                       most_certain_local_line_b = i;
+               }
+       }
+
+       /* No matches. */
+       if (most_certain_local_line_b == -1)
+               return;
+
+       most_certain_line_a = result[most_certain_local_line_b];
+
+       /*
+        * Subtract the most certain line's fingerprint in B from the matched
+        * fingerprint in A. This means that other lines in B can't also match
+        * the same parts of the line in A.
+        */
+       fingerprint_subtract(fingerprints_a + most_certain_line_a - start_a,
+                            fingerprints_b + most_certain_local_line_b);
+
+       /* Invalidate results that may be affected by the choice of most
+        * certain line.
+        */
+       invalidate_min = most_certain_local_line_b - max_search_distance_b;
+       invalidate_max = most_certain_local_line_b + max_search_distance_b + 1;
+       if (invalidate_min < 0)
+               invalidate_min = 0;
+       if (invalidate_max > length_b)
+               invalidate_max = length_b;
+
+       /* As the fingerprint in A has changed, discard previously calculated
+        * similarity values with that fingerprint.
+        */
+       for (i = invalidate_min; i < invalidate_max; ++i) {
+               closest_local_line_a = map_line_number(
+                       i + start_b, map_line_number_in_b_to_a) - start_a;
+
+               /* Check that the lines in A and B are close enough that there
+                * is a similarity value for them.
+                */
+               if (abs(most_certain_line_a - start_a - closest_local_line_a) >
+                       max_search_distance_a) {
+                       continue;
+               }
+
+               *get_similarity(similarities, most_certain_line_a - start_a,
+                               i, closest_local_line_a,
+                               max_search_distance_a) = -1;
+       }
+
+       /* More invalidating of results that may be affected by the choice of
+        * most certain line.
+        * Discard the matches for lines in B that are currently matched with a
+        * line in A such that their ordering contradicts the ordering imposed
+        * by the choice of most certain line.
+        */
+       for (i = most_certain_local_line_b - 1; i >= invalidate_min; --i) {
+               /* In this loop we discard results for lines in B that are
+                * before most-certain-line-B but are matched with a line in A
+                * that is after most-certain-line-A.
+                */
+               if (certainties[i] >= 0 &&
+                   (result[i] >= most_certain_line_a ||
+                    second_best_result[i] >= most_certain_line_a)) {
+                       certainties[i] = CERTAINTY_NOT_CALCULATED;
+               }
+       }
+       for (i = most_certain_local_line_b + 1; i < invalidate_max; ++i) {
+               /* In this loop we discard results for lines in B that are
+                * after most-certain-line-B but are matched with a line in A
+                * that is before most-certain-line-A.
+                */
+               if (certainties[i] >= 0 &&
+                   (result[i] <= most_certain_line_a ||
+                    second_best_result[i] <= most_certain_line_a)) {
+                       certainties[i] = CERTAINTY_NOT_CALCULATED;
+               }
+       }
+
+       /* Repeat the matching process for lines before the most certain line.
+        */
+       if (most_certain_local_line_b > 0) {
+               fuzzy_find_matching_lines_recurse(
+                       start_a, start_b,
+                       most_certain_line_a + 1 - start_a,
+                       most_certain_local_line_b,
+                       fingerprints_a, fingerprints_b, similarities,
+                       certainties, second_best_result, result,
+                       max_search_distance_a,
+                       max_search_distance_b,
+                       map_line_number_in_b_to_a);
+       }
+       /* Repeat the matching process for lines after the most certain line.
+        */
+       if (most_certain_local_line_b + 1 < length_b) {
+               second_half_start_a = most_certain_line_a;
+               offset_b = most_certain_local_line_b + 1;
+               second_half_start_b = start_b + offset_b;
+               second_half_length_a =
+                       length_a + start_a - second_half_start_a;
+               second_half_length_b =
+                       length_b + start_b - second_half_start_b;
+               fuzzy_find_matching_lines_recurse(
+                       second_half_start_a, second_half_start_b,
+                       second_half_length_a, second_half_length_b,
+                       fingerprints_a + second_half_start_a - start_a,
+                       fingerprints_b + offset_b,
+                       similarities +
+                               offset_b * (max_search_distance_a * 2 + 1),
+                       certainties + offset_b,
+                       second_best_result + offset_b, result + offset_b,
+                       max_search_distance_a,
+                       max_search_distance_b,
+                       map_line_number_in_b_to_a);
+       }
+}
+
+/* Find the lines in the parent line range that most closely match the lines in
+ * the target line range. This is accomplished by matching fingerprints in each
+ * blame_origin, and choosing the best matches that preserve the line ordering.
+ * See struct fingerprint for details of fingerprint matching, and
+ * fuzzy_find_matching_lines_recurse for details of preserving line ordering.
+ *
+ * The performance is believed to be O(n log n) in the typical case and O(n^2)
+ * in a pathological case, where n is the number of lines in the target range.
+ */
+static int *fuzzy_find_matching_lines(struct blame_origin *parent,
+                                     struct blame_origin *target,
+                                     int tlno, int parent_slno, int same,
+                                     int parent_len)
+{
+       /* We use the terminology "A" for the left hand side of the diff AKA
+        * parent, and "B" for the right hand side of the diff AKA target. */
+       int start_a = parent_slno;
+       int length_a = parent_len;
+       int start_b = tlno;
+       int length_b = same - tlno;
+
+       struct line_number_mapping map_line_number_in_b_to_a = {
+               start_a, length_a, start_b, length_b
+       };
+
+       struct fingerprint *fingerprints_a = parent->fingerprints;
+       struct fingerprint *fingerprints_b = target->fingerprints;
+
+       int i, *result, *second_best_result,
+               *certainties, *similarities, similarity_count;
+
+       /*
+        * max_search_distance_a means that given a line in B, compare it to
+        * the line in A that is closest to its position, and the lines in A
+        * that are no greater than max_search_distance_a lines away from the
+        * closest line in A.
+        *
+        * max_search_distance_b is an upper bound on the greatest possible
+        * distance between lines in B such that they will both be compared
+        * with the same line in A according to max_search_distance_a.
+        */
+       int max_search_distance_a = 10, max_search_distance_b;
+
+       if (length_a <= 0)
+               return NULL;
+
+       if (max_search_distance_a >= length_a)
+               max_search_distance_a = length_a ? length_a - 1 : 0;
+
+       max_search_distance_b = ((2 * max_search_distance_a + 1) * length_b
+                                - 1) / length_a;
+
+       result = xcalloc(sizeof(int), length_b);
+       second_best_result = xcalloc(sizeof(int), length_b);
+       certainties = xcalloc(sizeof(int), length_b);
+
+       /* See get_similarity() for details of similarities. */
+       similarity_count = length_b * (max_search_distance_a * 2 + 1);
+       similarities = xcalloc(sizeof(int), similarity_count);
+
+       for (i = 0; i < length_b; ++i) {
+               result[i] = -1;
+               second_best_result[i] = -1;
+               certainties[i] = CERTAINTY_NOT_CALCULATED;
+       }
+
+       for (i = 0; i < similarity_count; ++i)
+               similarities[i] = -1;
+
+       fuzzy_find_matching_lines_recurse(start_a, start_b,
+                                         length_a, length_b,
+                                         fingerprints_a + start_a,
+                                         fingerprints_b + start_b,
+                                         similarities,
+                                         certainties,
+                                         second_best_result,
+                                         result,
+                                         max_search_distance_a,
+                                         max_search_distance_b,
+                                         &map_line_number_in_b_to_a);
+
+       free(similarities);
+       free(certainties);
+       free(second_best_result);
+
+       return result;
+}
+
+static void fill_origin_fingerprints(struct blame_origin *o)
+{
+       int *line_starts;
+
+       if (o->fingerprints)
+               return;
+       o->num_lines = find_line_starts(&line_starts, o->file.ptr,
+                                       o->file.size);
+       o->fingerprints = xcalloc(sizeof(struct fingerprint), o->num_lines);
+       get_line_fingerprints(o->fingerprints, o->file.ptr, line_starts,
+                             0, o->num_lines);
+       free(line_starts);
+}
+
+static void drop_origin_fingerprints(struct blame_origin *o)
+{
+       if (o->fingerprints) {
+               free_line_fingerprints(o->fingerprints, o->num_lines);
+               o->num_lines = 0;
+               FREE_AND_NULL(o->fingerprints);
+       }
+}
+
 /*
  * Given an origin, prepare mmfile_t structure to be used by the
  * diff machinery
  */
 static void fill_origin_blob(struct diff_options *opt,
-                            struct blame_origin *o, mmfile_t *file, int *num_read_blob)
+                            struct blame_origin *o, mmfile_t *file,
+                            int *num_read_blob, int fill_fingerprints)
 {
        if (!o->file.ptr) {
                enum object_type type;
@@ -340,11 +1035,14 @@ static void fill_origin_blob(struct diff_options *opt,
        }
        else
                *file = o->file;
+       if (fill_fingerprints)
+               fill_origin_fingerprints(o);
 }
 
 static void drop_origin_blob(struct blame_origin *o)
 {
        FREE_AND_NULL(o->file.ptr);
+       drop_origin_fingerprints(o);
 }
 
 /*
@@ -480,7 +1178,9 @@ void blame_coalesce(struct blame_scoreboard *sb)
 
        for (ent = sb->ent; ent && (next = ent->next); ent = next) {
                if (ent->suspect == next->suspect &&
-                   ent->s_lno + ent->num_lines == next->s_lno) {
+                   ent->s_lno + ent->num_lines == next->s_lno &&
+                   ent->ignored == next->ignored &&
+                   ent->unblamable == next->unblamable) {
                        ent->num_lines += next->num_lines;
                        ent->next = next->next;
                        blame_origin_decref(next->suspect);
@@ -730,8 +1430,14 @@ static void split_overlap(struct blame_entry *split,
                          struct blame_origin *parent)
 {
        int chunk_end_lno;
+       int i;
        memset(split, 0, sizeof(struct blame_entry [3]));
 
+       for (i = 0; i < 3; i++) {
+               split[i].ignored = e->ignored;
+               split[i].unblamable = e->unblamable;
+       }
+
        if (e->s_lno < tlno) {
                /* there is a pre-chunk part not blamed on parent */
                split[0].suspect = blame_origin_incref(e->suspect);
@@ -839,6 +1545,164 @@ static struct blame_entry *reverse_blame(struct blame_entry *head,
        return tail;
 }
 
+/*
+ * Splits a blame entry into two entries at 'len' lines.  The original 'e'
+ * consists of len lines, i.e. [e->lno, e->lno + len), and the second part,
+ * which is returned, consists of the remainder: [e->lno + len, e->lno +
+ * e->num_lines).  The caller needs to sort out the reference counting for the
+ * new entry's suspect.
+ */
+static struct blame_entry *split_blame_at(struct blame_entry *e, int len,
+                                         struct blame_origin *new_suspect)
+{
+       struct blame_entry *n = xcalloc(1, sizeof(struct blame_entry));
+
+       n->suspect = new_suspect;
+       n->ignored = e->ignored;
+       n->unblamable = e->unblamable;
+       n->lno = e->lno + len;
+       n->s_lno = e->s_lno + len;
+       n->num_lines = e->num_lines - len;
+       e->num_lines = len;
+       e->score = 0;
+       return n;
+}
+
+struct blame_line_tracker {
+       int is_parent;
+       int s_lno;
+};
+
+static int are_lines_adjacent(struct blame_line_tracker *first,
+                             struct blame_line_tracker *second)
+{
+       return first->is_parent == second->is_parent &&
+              first->s_lno + 1 == second->s_lno;
+}
+
+static int scan_parent_range(struct fingerprint *p_fps,
+                            struct fingerprint *t_fps, int t_idx,
+                            int from, int nr_lines)
+{
+       int sim, p_idx;
+       #define FINGERPRINT_FILE_THRESHOLD      10
+       int best_sim_val = FINGERPRINT_FILE_THRESHOLD;
+       int best_sim_idx = -1;
+
+       for (p_idx = from; p_idx < from + nr_lines; p_idx++) {
+               sim = fingerprint_similarity(&t_fps[t_idx], &p_fps[p_idx]);
+               if (sim < best_sim_val)
+                       continue;
+               /* Break ties with the closest-to-target line number */
+               if (sim == best_sim_val && best_sim_idx != -1 &&
+                   abs(best_sim_idx - t_idx) < abs(p_idx - t_idx))
+                       continue;
+               best_sim_val = sim;
+               best_sim_idx = p_idx;
+       }
+       return best_sim_idx;
+}
+
+/*
+ * The first pass checks the blame entry (from the target) against the parent's
+ * diff chunk.  If that fails for a line, the second pass tries to match that
+ * line to any part of parent file.  That catches cases where a change was
+ * broken into two chunks by 'context.'
+ */
+static void guess_line_blames(struct blame_origin *parent,
+                             struct blame_origin *target,
+                             int tlno, int offset, int same, int parent_len,
+                             struct blame_line_tracker *line_blames)
+{
+       int i, best_idx, target_idx;
+       int parent_slno = tlno + offset;
+       int *fuzzy_matches;
+
+       fuzzy_matches = fuzzy_find_matching_lines(parent, target,
+                                                 tlno, parent_slno, same,
+                                                 parent_len);
+       for (i = 0; i < same - tlno; i++) {
+               target_idx = tlno + i;
+               if (fuzzy_matches && fuzzy_matches[i] >= 0) {
+                       best_idx = fuzzy_matches[i];
+               } else {
+                       best_idx = scan_parent_range(parent->fingerprints,
+                                                    target->fingerprints,
+                                                    target_idx, 0,
+                                                    parent->num_lines);
+               }
+               if (best_idx >= 0) {
+                       line_blames[i].is_parent = 1;
+                       line_blames[i].s_lno = best_idx;
+               } else {
+                       line_blames[i].is_parent = 0;
+                       line_blames[i].s_lno = target_idx;
+               }
+       }
+       free(fuzzy_matches);
+}
+
+/*
+ * This decides which parts of a blame entry go to the parent (added to the
+ * ignoredp list) and which stay with the target (added to the diffp list).  The
+ * actual decision was made in a separate heuristic function, and those answers
+ * for the lines in 'e' are in line_blames.  This consumes e, essentially
+ * putting it on a list.
+ *
+ * Note that the blame entries on the ignoredp list are not necessarily sorted
+ * with respect to the parent's line numbers yet.
+ */
+static void ignore_blame_entry(struct blame_entry *e,
+                              struct blame_origin *parent,
+                              struct blame_entry **diffp,
+                              struct blame_entry **ignoredp,
+                              struct blame_line_tracker *line_blames)
+{
+       int entry_len, nr_lines, i;
+
+       /*
+        * We carve new entries off the front of e.  Each entry comes from a
+        * contiguous chunk of lines: adjacent lines from the same origin
+        * (either the parent or the target).
+        */
+       entry_len = 1;
+       nr_lines = e->num_lines;        /* e changes in the loop */
+       for (i = 0; i < nr_lines; i++) {
+               struct blame_entry *next = NULL;
+
+               /*
+                * We are often adjacent to the next line - only split the blame
+                * entry when we have to.
+                */
+               if (i + 1 < nr_lines) {
+                       if (are_lines_adjacent(&line_blames[i],
+                                              &line_blames[i + 1])) {
+                               entry_len++;
+                               continue;
+                       }
+                       next = split_blame_at(e, entry_len,
+                                             blame_origin_incref(e->suspect));
+               }
+               if (line_blames[i].is_parent) {
+                       e->ignored = 1;
+                       blame_origin_decref(e->suspect);
+                       e->suspect = blame_origin_incref(parent);
+                       e->s_lno = line_blames[i - entry_len + 1].s_lno;
+                       e->next = *ignoredp;
+                       *ignoredp = e;
+               } else {
+                       e->unblamable = 1;
+                       /* e->s_lno is already in the target's address space. */
+                       e->next = *diffp;
+                       *diffp = e;
+               }
+               assert(e->num_lines == entry_len);
+               e = next;
+               entry_len = 1;
+       }
+       assert(!e);
+}
+
 /*
  * Process one hunk from the patch between the current suspect for
  * blame_entry e and its parent.  This first blames any unfinished
@@ -848,13 +1712,20 @@ static struct blame_entry *reverse_blame(struct blame_entry *head,
  * -C options may lead to overlapping/duplicate source line number
  * ranges, all we can rely on from sorting/merging is the order of the
  * first suspect line number.
+ *
+ * tlno: line number in the target where this chunk begins
+ * same: line number in the target where this chunk ends
+ * offset: add to tlno to get the chunk starting point in the parent
+ * parent_len: number of lines in the parent chunk
  */
 static void blame_chunk(struct blame_entry ***dstq, struct blame_entry ***srcq,
-                       int tlno, int offset, int same,
-                       struct blame_origin *parent)
+                       int tlno, int offset, int same, int parent_len,
+                       struct blame_origin *parent,
+                       struct blame_origin *target, int ignore_diffs)
 {
        struct blame_entry *e = **srcq;
-       struct blame_entry *samep = NULL, *diffp = NULL;
+       struct blame_entry *samep = NULL, *diffp = NULL, *ignoredp = NULL;
+       struct blame_line_tracker *line_blames = NULL;
 
        while (e && e->s_lno < tlno) {
                struct blame_entry *next = e->next;
@@ -865,14 +1736,9 @@ static void blame_chunk(struct blame_entry ***dstq, struct blame_entry ***srcq,
                 */
                if (e->s_lno + e->num_lines > tlno) {
                        /* Move second half to a new record */
-                       int len = tlno - e->s_lno;
-                       struct blame_entry *n = xcalloc(1, sizeof (struct blame_entry));
-                       n->suspect = e->suspect;
-                       n->lno = e->lno + len;
-                       n->s_lno = e->s_lno + len;
-                       n->num_lines = e->num_lines - len;
-                       e->num_lines = len;
-                       e->score = 0;
+                       struct blame_entry *n;
+
+                       n = split_blame_at(e, tlno - e->s_lno, e->suspect);
                        /* Push new record to diffp */
                        n->next = diffp;
                        diffp = n;
@@ -908,6 +1774,14 @@ static void blame_chunk(struct blame_entry ***dstq, struct blame_entry ***srcq,
         */
        samep = NULL;
        diffp = NULL;
+
+       if (ignore_diffs && same - tlno > 0) {
+               line_blames = xcalloc(sizeof(struct blame_line_tracker),
+                                     same - tlno);
+               guess_line_blames(parent, target, tlno, offset, same,
+                                 parent_len, line_blames);
+       }
+
        while (e && e->s_lno < same) {
                struct blame_entry *next = e->next;
 
@@ -919,22 +1793,37 @@ static void blame_chunk(struct blame_entry ***dstq, struct blame_entry ***srcq,
                         * Move second half to a new record to be
                         * processed by later chunks
                         */
-                       int len = same - e->s_lno;
-                       struct blame_entry *n = xcalloc(1, sizeof (struct blame_entry));
-                       n->suspect = blame_origin_incref(e->suspect);
-                       n->lno = e->lno + len;
-                       n->s_lno = e->s_lno + len;
-                       n->num_lines = e->num_lines - len;
-                       e->num_lines = len;
-                       e->score = 0;
+                       struct blame_entry *n;
+
+                       n = split_blame_at(e, same - e->s_lno,
+                                          blame_origin_incref(e->suspect));
                        /* Push new record to samep */
                        n->next = samep;
                        samep = n;
                }
-               e->next = diffp;
-               diffp = e;
+               if (ignore_diffs) {
+                       ignore_blame_entry(e, parent, &diffp, &ignoredp,
+                                          line_blames + e->s_lno - tlno);
+               } else {
+                       e->next = diffp;
+                       diffp = e;
+               }
                e = next;
        }
+       free(line_blames);
+       if (ignoredp) {
+               /*
+                * Note ignoredp is not sorted yet, and thus neither is dstq.
+                * That list must be sorted before we queue_blames().  We defer
+                * sorting until after all diff hunks are processed, so that
+                * guess_line_blames() can pick *any* line in the parent.  The
+                * slight drawback is that we end up sorting all blame entries
+                * passed to the parent, including those that are unrelated to
+                * changes made by the ignored commit.
+                */
+               **dstq = reverse_blame(ignoredp, **dstq);
+               *dstq = &ignoredp->next;
+       }
        **srcq = reverse_blame(diffp, reverse_blame(samep, e));
        /* Move across elements that are in the unblamable portion */
        if (diffp)
@@ -943,7 +1832,9 @@ static void blame_chunk(struct blame_entry ***dstq, struct blame_entry ***srcq,
 
 struct blame_chunk_cb_data {
        struct blame_origin *parent;
+       struct blame_origin *target;
        long offset;
+       int ignore_diffs;
        struct blame_entry **dstq;
        struct blame_entry **srcq;
 };
@@ -956,7 +1847,8 @@ static int blame_chunk_cb(long start_a, long count_a,
        if (start_a - start_b != d->offset)
                die("internal error in blame::blame_chunk_cb");
        blame_chunk(&d->dstq, &d->srcq, start_b, start_a - start_b,
-                   start_b + count_b, d->parent);
+                   start_b + count_b, count_a, d->parent, d->target,
+                   d->ignore_diffs);
        d->offset = start_a + count_a - (start_b + count_b);
        return 0;
 }
@@ -968,7 +1860,7 @@ static int blame_chunk_cb(long start_a, long count_a,
  */
 static void pass_blame_to_parent(struct blame_scoreboard *sb,
                                 struct blame_origin *target,
-                                struct blame_origin *parent)
+                                struct blame_origin *parent, int ignore_diffs)
 {
        mmfile_t file_p, file_o;
        struct blame_chunk_cb_data d;
@@ -978,11 +1870,15 @@ static void pass_blame_to_parent(struct blame_scoreboard *sb,
                return; /* nothing remains for this target */
 
        d.parent = parent;
+       d.target = target;
        d.offset = 0;
+       d.ignore_diffs = ignore_diffs;
        d.dstq = &newdest; d.srcq = &target->suspects;
 
-       fill_origin_blob(&sb->revs->diffopt, parent, &file_p, &sb->num_read_blob);
-       fill_origin_blob(&sb->revs->diffopt, target, &file_o, &sb->num_read_blob);
+       fill_origin_blob(&sb->revs->diffopt, parent, &file_p,
+                        &sb->num_read_blob, ignore_diffs);
+       fill_origin_blob(&sb->revs->diffopt, target, &file_o,
+                        &sb->num_read_blob, ignore_diffs);
        sb->num_get_patch++;
 
        if (diff_hunks(&file_p, &file_o, blame_chunk_cb, &d, sb->xdl_opts))
@@ -990,8 +1886,13 @@ static void pass_blame_to_parent(struct blame_scoreboard *sb,
                    oid_to_hex(&parent->commit->object.oid),
                    oid_to_hex(&target->commit->object.oid));
        /* The rest are the same as the parent */
-       blame_chunk(&d.dstq, &d.srcq, INT_MAX, d.offset, INT_MAX, parent);
+       blame_chunk(&d.dstq, &d.srcq, INT_MAX, d.offset, INT_MAX, 0,
+                   parent, target, 0);
        *d.dstq = NULL;
+       if (ignore_diffs)
+               newdest = llist_mergesort(newdest, get_next_blame,
+                                         set_next_blame,
+                                         compare_blame_suspect);
        queue_blames(sb, parent, newdest);
 
        return;
@@ -1188,7 +2089,8 @@ static void find_move_in_parent(struct blame_scoreboard *sb,
        if (!unblamed)
                return; /* nothing remains for this target */
 
-       fill_origin_blob(&sb->revs->diffopt, parent, &file_p, &sb->num_read_blob);
+       fill_origin_blob(&sb->revs->diffopt, parent, &file_p,
+                        &sb->num_read_blob, 0);
        if (!file_p.ptr)
                return;
 
@@ -1317,7 +2219,8 @@ static void find_copy_in_parent(struct blame_scoreboard *sb,
                        norigin = get_origin(parent, p->one->path);
                        oidcpy(&norigin->blob_oid, &p->one->oid);
                        norigin->mode = p->one->mode;
-                       fill_origin_blob(&sb->revs->diffopt, norigin, &file_p, &sb->num_read_blob);
+                       fill_origin_blob(&sb->revs->diffopt, norigin, &file_p,
+                                        &sb->num_read_blob, 0);
                        if (!file_p.ptr)
                                continue;
 
@@ -1495,11 +2398,34 @@ static void pass_blame(struct blame_scoreboard *sb, struct blame_origin *origin,
                        blame_origin_incref(porigin);
                        origin->previous = porigin;
                }
-               pass_blame_to_parent(sb, origin, porigin);
+               pass_blame_to_parent(sb, origin, porigin, 0);
                if (!origin->suspects)
                        goto finish;
        }
 
+       /*
+        * Pass remaining suspects for ignored commits to their parents.
+        */
+       if (oidset_contains(&sb->ignore_list, &commit->object.oid)) {
+               for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
+                    i < num_sg && sg;
+                    sg = sg->next, i++) {
+                       struct blame_origin *porigin = sg_origin[i];
+
+                       if (!porigin)
+                               continue;
+                       pass_blame_to_parent(sb, origin, porigin, 1);
+                       /*
+                        * Preemptively drop porigin so we can refresh the
+                        * fingerprints if we use the parent again, which can
+                        * occur if you ignore back-to-back commits.
+                        */
+                       drop_origin_blob(porigin);
+                       if (!origin->suspects)
+                               goto finish;
+               }
+       }
+
        /*
         * Optionally find moves in parents' files.
         */
@@ -1640,37 +2566,14 @@ void assign_blame(struct blame_scoreboard *sb, int opt)
        }
 }
 
-static const char *get_next_line(const char *start, const char *end)
-{
-       const char *nl = memchr(start, '\n', end - start);
-       return nl ? nl + 1 : end;
-}
-
 /*
  * To allow quick access to the contents of nth line in the
  * final image, prepare an index in the scoreboard.
  */
 static int prepare_lines(struct blame_scoreboard *sb)
 {
-       const char *buf = sb->final_buf;
-       unsigned long len = sb->final_buf_size;
-       const char *end = buf + len;
-       const char *p;
-       int *lineno;
-       int num = 0;
-
-       for (p = buf; p < end; p = get_next_line(p, end))
-               num++;
-
-       ALLOC_ARRAY(sb->lineno, num + 1);
-       lineno = sb->lineno;
-
-       for (p = buf; p < end; p = get_next_line(p, end))
-               *lineno++ = p - buf;
-
-       *lineno = len;
-
-       sb->num_lines = num;
+       sb->num_lines = find_line_starts(&sb->lineno, sb->final_buf,
+                                        sb->final_buf_size);
        return sb->num_lines;
 }