Merge branch 'zh/read-cache-copy-name-entry-fix'
[alt-git.git] / blame.c
blobda1052ac94bb47282fb6191e457dda6597c3e8fc
1 #include "cache.h"
2 #include "refs.h"
3 #include "object-store.h"
4 #include "cache-tree.h"
5 #include "mergesort.h"
6 #include "diff.h"
7 #include "diffcore.h"
8 #include "tag.h"
9 #include "blame.h"
10 #include "alloc.h"
11 #include "commit-slab.h"
12 #include "bloom.h"
13 #include "commit-graph.h"
15 define_commit_slab(blame_suspects, struct blame_origin *);
16 static struct blame_suspects blame_suspects;
18 struct blame_origin *get_blame_suspects(struct commit *commit)
20 struct blame_origin **result;
22 result = blame_suspects_peek(&blame_suspects, commit);
24 return result ? *result : NULL;
27 static void set_blame_suspects(struct commit *commit, struct blame_origin *origin)
29 *blame_suspects_at(&blame_suspects, commit) = origin;
32 void blame_origin_decref(struct blame_origin *o)
34 if (o && --o->refcnt <= 0) {
35 struct blame_origin *p, *l = NULL;
36 if (o->previous)
37 blame_origin_decref(o->previous);
38 free(o->file.ptr);
39 /* Should be present exactly once in commit chain */
40 for (p = get_blame_suspects(o->commit); p; l = p, p = p->next) {
41 if (p == o) {
42 if (l)
43 l->next = p->next;
44 else
45 set_blame_suspects(o->commit, p->next);
46 free(o);
47 return;
50 die("internal error in blame_origin_decref");
55 * Given a commit and a path in it, create a new origin structure.
56 * The callers that add blame to the scoreboard should use
57 * get_origin() to obtain shared, refcounted copy instead of calling
58 * this function directly.
60 static struct blame_origin *make_origin(struct commit *commit, const char *path)
62 struct blame_origin *o;
63 FLEX_ALLOC_STR(o, path, path);
64 o->commit = commit;
65 o->refcnt = 1;
66 o->next = get_blame_suspects(commit);
67 set_blame_suspects(commit, o);
68 return o;
72 * Locate an existing origin or create a new one.
73 * This moves the origin to front position in the commit util list.
75 static struct blame_origin *get_origin(struct commit *commit, const char *path)
77 struct blame_origin *o, *l;
79 for (o = get_blame_suspects(commit), l = NULL; o; l = o, o = o->next) {
80 if (!strcmp(o->path, path)) {
81 /* bump to front */
82 if (l) {
83 l->next = o->next;
84 o->next = get_blame_suspects(commit);
85 set_blame_suspects(commit, o);
87 return blame_origin_incref(o);
90 return make_origin(commit, path);
95 static void verify_working_tree_path(struct repository *r,
96 struct commit *work_tree, const char *path)
98 struct commit_list *parents;
99 int pos;
101 for (parents = work_tree->parents; parents; parents = parents->next) {
102 const struct object_id *commit_oid = &parents->item->object.oid;
103 struct object_id blob_oid;
104 unsigned short mode;
106 if (!get_tree_entry(r, commit_oid, path, &blob_oid, &mode) &&
107 oid_object_info(r, &blob_oid, NULL) == OBJ_BLOB)
108 return;
111 pos = index_name_pos(r->index, path, strlen(path));
112 if (pos >= 0)
113 ; /* path is in the index */
114 else if (-1 - pos < r->index->cache_nr &&
115 !strcmp(r->index->cache[-1 - pos]->name, path))
116 ; /* path is in the index, unmerged */
117 else
118 die("no such path '%s' in HEAD", path);
121 static struct commit_list **append_parent(struct repository *r,
122 struct commit_list **tail,
123 const struct object_id *oid)
125 struct commit *parent;
127 parent = lookup_commit_reference(r, oid);
128 if (!parent)
129 die("no such commit %s", oid_to_hex(oid));
130 return &commit_list_insert(parent, tail)->next;
133 static void append_merge_parents(struct repository *r,
134 struct commit_list **tail)
136 int merge_head;
137 struct strbuf line = STRBUF_INIT;
139 merge_head = open(git_path_merge_head(r), O_RDONLY);
140 if (merge_head < 0) {
141 if (errno == ENOENT)
142 return;
143 die("cannot open '%s' for reading",
144 git_path_merge_head(r));
147 while (!strbuf_getwholeline_fd(&line, merge_head, '\n')) {
148 struct object_id oid;
149 if (get_oid_hex(line.buf, &oid))
150 die("unknown line in '%s': %s",
151 git_path_merge_head(r), line.buf);
152 tail = append_parent(r, tail, &oid);
154 close(merge_head);
155 strbuf_release(&line);
159 * This isn't as simple as passing sb->buf and sb->len, because we
160 * want to transfer ownership of the buffer to the commit (so we
161 * must use detach).
163 static void set_commit_buffer_from_strbuf(struct repository *r,
164 struct commit *c,
165 struct strbuf *sb)
167 size_t len;
168 void *buf = strbuf_detach(sb, &len);
169 set_commit_buffer(r, c, buf, len);
173 * Prepare a dummy commit that represents the work tree (or staged) item.
174 * Note that annotating work tree item never works in the reverse.
176 static struct commit *fake_working_tree_commit(struct repository *r,
177 struct diff_options *opt,
178 const char *path,
179 const char *contents_from)
181 struct commit *commit;
182 struct blame_origin *origin;
183 struct commit_list **parent_tail, *parent;
184 struct object_id head_oid;
185 struct strbuf buf = STRBUF_INIT;
186 const char *ident;
187 time_t now;
188 int len;
189 struct cache_entry *ce;
190 unsigned mode;
191 struct strbuf msg = STRBUF_INIT;
193 repo_read_index(r);
194 time(&now);
195 commit = alloc_commit_node(r);
196 commit->object.parsed = 1;
197 commit->date = now;
198 parent_tail = &commit->parents;
200 if (!resolve_ref_unsafe("HEAD", RESOLVE_REF_READING, &head_oid, NULL))
201 die("no such ref: HEAD");
203 parent_tail = append_parent(r, parent_tail, &head_oid);
204 append_merge_parents(r, parent_tail);
205 verify_working_tree_path(r, commit, path);
207 origin = make_origin(commit, path);
209 ident = fmt_ident("Not Committed Yet", "not.committed.yet",
210 WANT_BLANK_IDENT, NULL, 0);
211 strbuf_addstr(&msg, "tree 0000000000000000000000000000000000000000\n");
212 for (parent = commit->parents; parent; parent = parent->next)
213 strbuf_addf(&msg, "parent %s\n",
214 oid_to_hex(&parent->item->object.oid));
215 strbuf_addf(&msg,
216 "author %s\n"
217 "committer %s\n\n"
218 "Version of %s from %s\n",
219 ident, ident, path,
220 (!contents_from ? path :
221 (!strcmp(contents_from, "-") ? "standard input" : contents_from)));
222 set_commit_buffer_from_strbuf(r, commit, &msg);
224 if (!contents_from || strcmp("-", contents_from)) {
225 struct stat st;
226 const char *read_from;
227 char *buf_ptr;
228 unsigned long buf_len;
230 if (contents_from) {
231 if (stat(contents_from, &st) < 0)
232 die_errno("Cannot stat '%s'", contents_from);
233 read_from = contents_from;
235 else {
236 if (lstat(path, &st) < 0)
237 die_errno("Cannot lstat '%s'", path);
238 read_from = path;
240 mode = canon_mode(st.st_mode);
242 switch (st.st_mode & S_IFMT) {
243 case S_IFREG:
244 if (opt->flags.allow_textconv &&
245 textconv_object(r, read_from, mode, null_oid(), 0, &buf_ptr, &buf_len))
246 strbuf_attach(&buf, buf_ptr, buf_len, buf_len + 1);
247 else if (strbuf_read_file(&buf, read_from, st.st_size) != st.st_size)
248 die_errno("cannot open or read '%s'", read_from);
249 break;
250 case S_IFLNK:
251 if (strbuf_readlink(&buf, read_from, st.st_size) < 0)
252 die_errno("cannot readlink '%s'", read_from);
253 break;
254 default:
255 die("unsupported file type %s", read_from);
258 else {
259 /* Reading from stdin */
260 mode = 0;
261 if (strbuf_read(&buf, 0, 0) < 0)
262 die_errno("failed to read from stdin");
264 convert_to_git(r->index, path, buf.buf, buf.len, &buf, 0);
265 origin->file.ptr = buf.buf;
266 origin->file.size = buf.len;
267 pretend_object_file(buf.buf, buf.len, OBJ_BLOB, &origin->blob_oid);
270 * Read the current index, replace the path entry with
271 * origin->blob_sha1 without mucking with its mode or type
272 * bits; we are not going to write this index out -- we just
273 * want to run "diff-index --cached".
275 discard_index(r->index);
276 repo_read_index(r);
278 len = strlen(path);
279 if (!mode) {
280 int pos = index_name_pos(r->index, path, len);
281 if (0 <= pos)
282 mode = r->index->cache[pos]->ce_mode;
283 else
284 /* Let's not bother reading from HEAD tree */
285 mode = S_IFREG | 0644;
287 ce = make_empty_cache_entry(r->index, len);
288 oidcpy(&ce->oid, &origin->blob_oid);
289 memcpy(ce->name, path, len);
290 ce->ce_flags = create_ce_flags(0);
291 ce->ce_namelen = len;
292 ce->ce_mode = create_ce_mode(mode);
293 add_index_entry(r->index, ce,
294 ADD_CACHE_OK_TO_ADD | ADD_CACHE_OK_TO_REPLACE);
296 cache_tree_invalidate_path(r->index, path);
298 return commit;
303 static int diff_hunks(mmfile_t *file_a, mmfile_t *file_b,
304 xdl_emit_hunk_consume_func_t hunk_func, void *cb_data, int xdl_opts)
306 xpparam_t xpp = {0};
307 xdemitconf_t xecfg = {0};
308 xdemitcb_t ecb = {NULL};
310 xpp.flags = xdl_opts;
311 xecfg.hunk_func = hunk_func;
312 ecb.priv = cb_data;
313 return xdi_diff(file_a, file_b, &xpp, &xecfg, &ecb);
316 static const char *get_next_line(const char *start, const char *end)
318 const char *nl = memchr(start, '\n', end - start);
320 return nl ? nl + 1 : end;
323 static int find_line_starts(int **line_starts, const char *buf,
324 unsigned long len)
326 const char *end = buf + len;
327 const char *p;
328 int *lineno;
329 int num = 0;
331 for (p = buf; p < end; p = get_next_line(p, end))
332 num++;
334 ALLOC_ARRAY(*line_starts, num + 1);
335 lineno = *line_starts;
337 for (p = buf; p < end; p = get_next_line(p, end))
338 *lineno++ = p - buf;
340 *lineno = len;
342 return num;
345 struct fingerprint_entry;
347 /* A fingerprint is intended to loosely represent a string, such that two
348 * fingerprints can be quickly compared to give an indication of the similarity
349 * of the strings that they represent.
351 * A fingerprint is represented as a multiset of the lower-cased byte pairs in
352 * the string that it represents. Whitespace is added at each end of the
353 * string. Whitespace pairs are ignored. Whitespace is converted to '\0'.
354 * For example, the string "Darth Radar" will be converted to the following
355 * fingerprint:
356 * {"\0d", "da", "da", "ar", "ar", "rt", "th", "h\0", "\0r", "ra", "ad", "r\0"}
358 * The similarity between two fingerprints is the size of the intersection of
359 * their multisets, including repeated elements. See fingerprint_similarity for
360 * examples.
362 * For ease of implementation, the fingerprint is implemented as a map
363 * of byte pairs to the count of that byte pair in the string, instead of
364 * allowing repeated elements in a set.
366 struct fingerprint {
367 struct hashmap map;
368 /* As we know the maximum number of entries in advance, it's
369 * convenient to store the entries in a single array instead of having
370 * the hashmap manage the memory.
372 struct fingerprint_entry *entries;
375 /* A byte pair in a fingerprint. Stores the number of times the byte pair
376 * occurs in the string that the fingerprint represents.
378 struct fingerprint_entry {
379 /* The hashmap entry - the hash represents the byte pair in its
380 * entirety so we don't need to store the byte pair separately.
382 struct hashmap_entry entry;
383 /* The number of times the byte pair occurs in the string that the
384 * fingerprint represents.
386 int count;
389 /* See `struct fingerprint` for an explanation of what a fingerprint is.
390 * \param result the fingerprint of the string is stored here. This must be
391 * freed later using free_fingerprint.
392 * \param line_begin the start of the string
393 * \param line_end the end of the string
395 static void get_fingerprint(struct fingerprint *result,
396 const char *line_begin,
397 const char *line_end)
399 unsigned int hash, c0 = 0, c1;
400 const char *p;
401 int max_map_entry_count = 1 + line_end - line_begin;
402 struct fingerprint_entry *entry = xcalloc(max_map_entry_count,
403 sizeof(struct fingerprint_entry));
404 struct fingerprint_entry *found_entry;
406 hashmap_init(&result->map, NULL, NULL, max_map_entry_count);
407 result->entries = entry;
408 for (p = line_begin; p <= line_end; ++p, c0 = c1) {
409 /* Always terminate the string with whitespace.
410 * Normalise whitespace to 0, and normalise letters to
411 * lower case. This won't work for multibyte characters but at
412 * worst will match some unrelated characters.
414 if ((p == line_end) || isspace(*p))
415 c1 = 0;
416 else
417 c1 = tolower(*p);
418 hash = c0 | (c1 << 8);
419 /* Ignore whitespace pairs */
420 if (hash == 0)
421 continue;
422 hashmap_entry_init(&entry->entry, hash);
424 found_entry = hashmap_get_entry(&result->map, entry,
425 /* member name */ entry, NULL);
426 if (found_entry) {
427 found_entry->count += 1;
428 } else {
429 entry->count = 1;
430 hashmap_add(&result->map, &entry->entry);
431 ++entry;
436 static void free_fingerprint(struct fingerprint *f)
438 hashmap_clear(&f->map);
439 free(f->entries);
442 /* Calculates the similarity between two fingerprints as the size of the
443 * intersection of their multisets, including repeated elements. See
444 * `struct fingerprint` for an explanation of the fingerprint representation.
445 * The similarity between "cat mat" and "father rather" is 2 because "at" is
446 * present twice in both strings while the similarity between "tim" and "mit"
447 * is 0.
449 static int fingerprint_similarity(struct fingerprint *a, struct fingerprint *b)
451 int intersection = 0;
452 struct hashmap_iter iter;
453 const struct fingerprint_entry *entry_a, *entry_b;
455 hashmap_for_each_entry(&b->map, &iter, entry_b,
456 entry /* member name */) {
457 entry_a = hashmap_get_entry(&a->map, entry_b, entry, NULL);
458 if (entry_a) {
459 intersection += entry_a->count < entry_b->count ?
460 entry_a->count : entry_b->count;
463 return intersection;
466 /* Subtracts byte-pair elements in B from A, modifying A in place.
468 static void fingerprint_subtract(struct fingerprint *a, struct fingerprint *b)
470 struct hashmap_iter iter;
471 struct fingerprint_entry *entry_a;
472 const struct fingerprint_entry *entry_b;
474 hashmap_iter_init(&b->map, &iter);
476 hashmap_for_each_entry(&b->map, &iter, entry_b,
477 entry /* member name */) {
478 entry_a = hashmap_get_entry(&a->map, entry_b, entry, NULL);
479 if (entry_a) {
480 if (entry_a->count <= entry_b->count)
481 hashmap_remove(&a->map, &entry_b->entry, NULL);
482 else
483 entry_a->count -= entry_b->count;
488 /* Calculate fingerprints for a series of lines.
489 * Puts the fingerprints in the fingerprints array, which must have been
490 * preallocated to allow storing line_count elements.
492 static void get_line_fingerprints(struct fingerprint *fingerprints,
493 const char *content, const int *line_starts,
494 long first_line, long line_count)
496 int i;
497 const char *linestart, *lineend;
499 line_starts += first_line;
500 for (i = 0; i < line_count; ++i) {
501 linestart = content + line_starts[i];
502 lineend = content + line_starts[i + 1];
503 get_fingerprint(fingerprints + i, linestart, lineend);
507 static void free_line_fingerprints(struct fingerprint *fingerprints,
508 int nr_fingerprints)
510 int i;
512 for (i = 0; i < nr_fingerprints; i++)
513 free_fingerprint(&fingerprints[i]);
516 /* This contains the data necessary to linearly map a line number in one half
517 * of a diff chunk to the line in the other half of the diff chunk that is
518 * closest in terms of its position as a fraction of the length of the chunk.
520 struct line_number_mapping {
521 int destination_start, destination_length,
522 source_start, source_length;
525 /* Given a line number in one range, offset and scale it to map it onto the
526 * other range.
527 * Essentially this mapping is a simple linear equation but the calculation is
528 * more complicated to allow performing it with integer operations.
529 * Another complication is that if a line could map onto many lines in the
530 * destination range then we want to choose the line at the center of those
531 * possibilities.
532 * Example: if the chunk is 2 lines long in A and 10 lines long in B then the
533 * first 5 lines in B will map onto the first line in the A chunk, while the
534 * last 5 lines will all map onto the second line in the A chunk.
535 * Example: if the chunk is 10 lines long in A and 2 lines long in B then line
536 * 0 in B will map onto line 2 in A, and line 1 in B will map onto line 7 in A.
538 static int map_line_number(int line_number,
539 const struct line_number_mapping *mapping)
541 return ((line_number - mapping->source_start) * 2 + 1) *
542 mapping->destination_length /
543 (mapping->source_length * 2) +
544 mapping->destination_start;
547 /* Get a pointer to the element storing the similarity between a line in A
548 * and a line in B.
550 * The similarities are stored in a 2-dimensional array. Each "row" in the
551 * array contains the similarities for a line in B. The similarities stored in
552 * a row are the similarities between the line in B and the nearby lines in A.
553 * To keep the length of each row the same, it is padded out with values of -1
554 * where the search range extends beyond the lines in A.
555 * For example, if max_search_distance_a is 2 and the two sides of a diff chunk
556 * look like this:
557 * a | m
558 * b | n
559 * c | o
560 * d | p
561 * e | q
562 * Then the similarity array will contain:
563 * [-1, -1, am, bm, cm,
564 * -1, an, bn, cn, dn,
565 * ao, bo, co, do, eo,
566 * bp, cp, dp, ep, -1,
567 * cq, dq, eq, -1, -1]
568 * Where similarities are denoted either by -1 for invalid, or the
569 * concatenation of the two lines in the diff being compared.
571 * \param similarities array of similarities between lines in A and B
572 * \param line_a the index of the line in A, in the same frame of reference as
573 * closest_line_a.
574 * \param local_line_b the index of the line in B, relative to the first line
575 * in B that similarities represents.
576 * \param closest_line_a the index of the line in A that is deemed to be
577 * closest to local_line_b. This must be in the same
578 * frame of reference as line_a. This value defines
579 * where similarities is centered for the line in B.
580 * \param max_search_distance_a maximum distance in lines from the closest line
581 * in A for other lines in A for which
582 * similarities may be calculated.
584 static int *get_similarity(int *similarities,
585 int line_a, int local_line_b,
586 int closest_line_a, int max_search_distance_a)
588 assert(abs(line_a - closest_line_a) <=
589 max_search_distance_a);
590 return similarities + line_a - closest_line_a +
591 max_search_distance_a +
592 local_line_b * (max_search_distance_a * 2 + 1);
595 #define CERTAIN_NOTHING_MATCHES -2
596 #define CERTAINTY_NOT_CALCULATED -1
598 /* Given a line in B, first calculate its similarities with nearby lines in A
599 * if not already calculated, then identify the most similar and second most
600 * similar lines. The "certainty" is calculated based on those two
601 * similarities.
603 * \param start_a the index of the first line of the chunk in A
604 * \param length_a the length in lines of the chunk in A
605 * \param local_line_b the index of the line in B, relative to the first line
606 * in the chunk.
607 * \param fingerprints_a array of fingerprints for the chunk in A
608 * \param fingerprints_b array of fingerprints for the chunk in B
609 * \param similarities 2-dimensional array of similarities between lines in A
610 * and B. See get_similarity() for more details.
611 * \param certainties array of values indicating how strongly a line in B is
612 * matched with some line in A.
613 * \param second_best_result array of absolute indices in A for the second
614 * closest match of a line in B.
615 * \param result array of absolute indices in A for the closest match of a line
616 * in B.
617 * \param max_search_distance_a maximum distance in lines from the closest line
618 * in A for other lines in A for which
619 * similarities may be calculated.
620 * \param map_line_number_in_b_to_a parameter to map_line_number().
622 static void find_best_line_matches(
623 int start_a,
624 int length_a,
625 int start_b,
626 int local_line_b,
627 struct fingerprint *fingerprints_a,
628 struct fingerprint *fingerprints_b,
629 int *similarities,
630 int *certainties,
631 int *second_best_result,
632 int *result,
633 const int max_search_distance_a,
634 const struct line_number_mapping *map_line_number_in_b_to_a)
637 int i, search_start, search_end, closest_local_line_a, *similarity,
638 best_similarity = 0, second_best_similarity = 0,
639 best_similarity_index = 0, second_best_similarity_index = 0;
641 /* certainty has already been calculated so no need to redo the work */
642 if (certainties[local_line_b] != CERTAINTY_NOT_CALCULATED)
643 return;
645 closest_local_line_a = map_line_number(
646 local_line_b + start_b, map_line_number_in_b_to_a) - start_a;
648 search_start = closest_local_line_a - max_search_distance_a;
649 if (search_start < 0)
650 search_start = 0;
652 search_end = closest_local_line_a + max_search_distance_a + 1;
653 if (search_end > length_a)
654 search_end = length_a;
656 for (i = search_start; i < search_end; ++i) {
657 similarity = get_similarity(similarities,
658 i, local_line_b,
659 closest_local_line_a,
660 max_search_distance_a);
661 if (*similarity == -1) {
662 /* This value will never exceed 10 but assert just in
663 * case
665 assert(abs(i - closest_local_line_a) < 1000);
666 /* scale the similarity by (1000 - distance from
667 * closest line) to act as a tie break between lines
668 * that otherwise are equally similar.
670 *similarity = fingerprint_similarity(
671 fingerprints_b + local_line_b,
672 fingerprints_a + i) *
673 (1000 - abs(i - closest_local_line_a));
675 if (*similarity > best_similarity) {
676 second_best_similarity = best_similarity;
677 second_best_similarity_index = best_similarity_index;
678 best_similarity = *similarity;
679 best_similarity_index = i;
680 } else if (*similarity > second_best_similarity) {
681 second_best_similarity = *similarity;
682 second_best_similarity_index = i;
686 if (best_similarity == 0) {
687 /* this line definitely doesn't match with anything. Mark it
688 * with this special value so it doesn't get invalidated and
689 * won't be recalculated.
691 certainties[local_line_b] = CERTAIN_NOTHING_MATCHES;
692 result[local_line_b] = -1;
693 } else {
694 /* Calculate the certainty with which this line matches.
695 * If the line matches well with two lines then that reduces
696 * the certainty. However we still want to prioritise matching
697 * a line that matches very well with two lines over matching a
698 * line that matches poorly with one line, hence doubling
699 * best_similarity.
700 * This means that if we have
701 * line X that matches only one line with a score of 3,
702 * line Y that matches two lines equally with a score of 5,
703 * and line Z that matches only one line with a score or 2,
704 * then the lines in order of certainty are X, Y, Z.
706 certainties[local_line_b] = best_similarity * 2 -
707 second_best_similarity;
709 /* We keep both the best and second best results to allow us to
710 * check at a later stage of the matching process whether the
711 * result needs to be invalidated.
713 result[local_line_b] = start_a + best_similarity_index;
714 second_best_result[local_line_b] =
715 start_a + second_best_similarity_index;
720 * This finds the line that we can match with the most confidence, and
721 * uses it as a partition. It then calls itself on the lines on either side of
722 * that partition. In this way we avoid lines appearing out of order, and
723 * retain a sensible line ordering.
724 * \param start_a index of the first line in A with which lines in B may be
725 * compared.
726 * \param start_b index of the first line in B for which matching should be
727 * done.
728 * \param length_a number of lines in A with which lines in B may be compared.
729 * \param length_b number of lines in B for which matching should be done.
730 * \param fingerprints_a mutable array of fingerprints in A. The first element
731 * corresponds to the line at start_a.
732 * \param fingerprints_b array of fingerprints in B. The first element
733 * corresponds to the line at start_b.
734 * \param similarities 2-dimensional array of similarities between lines in A
735 * and B. See get_similarity() for more details.
736 * \param certainties array of values indicating how strongly a line in B is
737 * matched with some line in A.
738 * \param second_best_result array of absolute indices in A for the second
739 * closest match of a line in B.
740 * \param result array of absolute indices in A for the closest match of a line
741 * in B.
742 * \param max_search_distance_a maximum distance in lines from the closest line
743 * in A for other lines in A for which
744 * similarities may be calculated.
745 * \param max_search_distance_b an upper bound on the greatest possible
746 * distance between lines in B such that they will
747 * both be compared with the same line in A
748 * according to max_search_distance_a.
749 * \param map_line_number_in_b_to_a parameter to map_line_number().
751 static void fuzzy_find_matching_lines_recurse(
752 int start_a, int start_b,
753 int length_a, int length_b,
754 struct fingerprint *fingerprints_a,
755 struct fingerprint *fingerprints_b,
756 int *similarities,
757 int *certainties,
758 int *second_best_result,
759 int *result,
760 int max_search_distance_a,
761 int max_search_distance_b,
762 const struct line_number_mapping *map_line_number_in_b_to_a)
764 int i, invalidate_min, invalidate_max, offset_b,
765 second_half_start_a, second_half_start_b,
766 second_half_length_a, second_half_length_b,
767 most_certain_line_a, most_certain_local_line_b = -1,
768 most_certain_line_certainty = -1,
769 closest_local_line_a;
771 for (i = 0; i < length_b; ++i) {
772 find_best_line_matches(start_a,
773 length_a,
774 start_b,
776 fingerprints_a,
777 fingerprints_b,
778 similarities,
779 certainties,
780 second_best_result,
781 result,
782 max_search_distance_a,
783 map_line_number_in_b_to_a);
785 if (certainties[i] > most_certain_line_certainty) {
786 most_certain_line_certainty = certainties[i];
787 most_certain_local_line_b = i;
791 /* No matches. */
792 if (most_certain_local_line_b == -1)
793 return;
795 most_certain_line_a = result[most_certain_local_line_b];
798 * Subtract the most certain line's fingerprint in B from the matched
799 * fingerprint in A. This means that other lines in B can't also match
800 * the same parts of the line in A.
802 fingerprint_subtract(fingerprints_a + most_certain_line_a - start_a,
803 fingerprints_b + most_certain_local_line_b);
805 /* Invalidate results that may be affected by the choice of most
806 * certain line.
808 invalidate_min = most_certain_local_line_b - max_search_distance_b;
809 invalidate_max = most_certain_local_line_b + max_search_distance_b + 1;
810 if (invalidate_min < 0)
811 invalidate_min = 0;
812 if (invalidate_max > length_b)
813 invalidate_max = length_b;
815 /* As the fingerprint in A has changed, discard previously calculated
816 * similarity values with that fingerprint.
818 for (i = invalidate_min; i < invalidate_max; ++i) {
819 closest_local_line_a = map_line_number(
820 i + start_b, map_line_number_in_b_to_a) - start_a;
822 /* Check that the lines in A and B are close enough that there
823 * is a similarity value for them.
825 if (abs(most_certain_line_a - start_a - closest_local_line_a) >
826 max_search_distance_a) {
827 continue;
830 *get_similarity(similarities, most_certain_line_a - start_a,
831 i, closest_local_line_a,
832 max_search_distance_a) = -1;
835 /* More invalidating of results that may be affected by the choice of
836 * most certain line.
837 * Discard the matches for lines in B that are currently matched with a
838 * line in A such that their ordering contradicts the ordering imposed
839 * by the choice of most certain line.
841 for (i = most_certain_local_line_b - 1; i >= invalidate_min; --i) {
842 /* In this loop we discard results for lines in B that are
843 * before most-certain-line-B but are matched with a line in A
844 * that is after most-certain-line-A.
846 if (certainties[i] >= 0 &&
847 (result[i] >= most_certain_line_a ||
848 second_best_result[i] >= most_certain_line_a)) {
849 certainties[i] = CERTAINTY_NOT_CALCULATED;
852 for (i = most_certain_local_line_b + 1; i < invalidate_max; ++i) {
853 /* In this loop we discard results for lines in B that are
854 * after most-certain-line-B but are matched with a line in A
855 * that is before most-certain-line-A.
857 if (certainties[i] >= 0 &&
858 (result[i] <= most_certain_line_a ||
859 second_best_result[i] <= most_certain_line_a)) {
860 certainties[i] = CERTAINTY_NOT_CALCULATED;
864 /* Repeat the matching process for lines before the most certain line.
866 if (most_certain_local_line_b > 0) {
867 fuzzy_find_matching_lines_recurse(
868 start_a, start_b,
869 most_certain_line_a + 1 - start_a,
870 most_certain_local_line_b,
871 fingerprints_a, fingerprints_b, similarities,
872 certainties, second_best_result, result,
873 max_search_distance_a,
874 max_search_distance_b,
875 map_line_number_in_b_to_a);
877 /* Repeat the matching process for lines after the most certain line.
879 if (most_certain_local_line_b + 1 < length_b) {
880 second_half_start_a = most_certain_line_a;
881 offset_b = most_certain_local_line_b + 1;
882 second_half_start_b = start_b + offset_b;
883 second_half_length_a =
884 length_a + start_a - second_half_start_a;
885 second_half_length_b =
886 length_b + start_b - second_half_start_b;
887 fuzzy_find_matching_lines_recurse(
888 second_half_start_a, second_half_start_b,
889 second_half_length_a, second_half_length_b,
890 fingerprints_a + second_half_start_a - start_a,
891 fingerprints_b + offset_b,
892 similarities +
893 offset_b * (max_search_distance_a * 2 + 1),
894 certainties + offset_b,
895 second_best_result + offset_b, result + offset_b,
896 max_search_distance_a,
897 max_search_distance_b,
898 map_line_number_in_b_to_a);
902 /* Find the lines in the parent line range that most closely match the lines in
903 * the target line range. This is accomplished by matching fingerprints in each
904 * blame_origin, and choosing the best matches that preserve the line ordering.
905 * See struct fingerprint for details of fingerprint matching, and
906 * fuzzy_find_matching_lines_recurse for details of preserving line ordering.
908 * The performance is believed to be O(n log n) in the typical case and O(n^2)
909 * in a pathological case, where n is the number of lines in the target range.
911 static int *fuzzy_find_matching_lines(struct blame_origin *parent,
912 struct blame_origin *target,
913 int tlno, int parent_slno, int same,
914 int parent_len)
916 /* We use the terminology "A" for the left hand side of the diff AKA
917 * parent, and "B" for the right hand side of the diff AKA target. */
918 int start_a = parent_slno;
919 int length_a = parent_len;
920 int start_b = tlno;
921 int length_b = same - tlno;
923 struct line_number_mapping map_line_number_in_b_to_a = {
924 start_a, length_a, start_b, length_b
927 struct fingerprint *fingerprints_a = parent->fingerprints;
928 struct fingerprint *fingerprints_b = target->fingerprints;
930 int i, *result, *second_best_result,
931 *certainties, *similarities, similarity_count;
934 * max_search_distance_a means that given a line in B, compare it to
935 * the line in A that is closest to its position, and the lines in A
936 * that are no greater than max_search_distance_a lines away from the
937 * closest line in A.
939 * max_search_distance_b is an upper bound on the greatest possible
940 * distance between lines in B such that they will both be compared
941 * with the same line in A according to max_search_distance_a.
943 int max_search_distance_a = 10, max_search_distance_b;
945 if (length_a <= 0)
946 return NULL;
948 if (max_search_distance_a >= length_a)
949 max_search_distance_a = length_a ? length_a - 1 : 0;
951 max_search_distance_b = ((2 * max_search_distance_a + 1) * length_b
952 - 1) / length_a;
954 CALLOC_ARRAY(result, length_b);
955 CALLOC_ARRAY(second_best_result, length_b);
956 CALLOC_ARRAY(certainties, length_b);
958 /* See get_similarity() for details of similarities. */
959 similarity_count = length_b * (max_search_distance_a * 2 + 1);
960 CALLOC_ARRAY(similarities, similarity_count);
962 for (i = 0; i < length_b; ++i) {
963 result[i] = -1;
964 second_best_result[i] = -1;
965 certainties[i] = CERTAINTY_NOT_CALCULATED;
968 for (i = 0; i < similarity_count; ++i)
969 similarities[i] = -1;
971 fuzzy_find_matching_lines_recurse(start_a, start_b,
972 length_a, length_b,
973 fingerprints_a + start_a,
974 fingerprints_b + start_b,
975 similarities,
976 certainties,
977 second_best_result,
978 result,
979 max_search_distance_a,
980 max_search_distance_b,
981 &map_line_number_in_b_to_a);
983 free(similarities);
984 free(certainties);
985 free(second_best_result);
987 return result;
990 static void fill_origin_fingerprints(struct blame_origin *o)
992 int *line_starts;
994 if (o->fingerprints)
995 return;
996 o->num_lines = find_line_starts(&line_starts, o->file.ptr,
997 o->file.size);
998 CALLOC_ARRAY(o->fingerprints, o->num_lines);
999 get_line_fingerprints(o->fingerprints, o->file.ptr, line_starts,
1000 0, o->num_lines);
1001 free(line_starts);
1004 static void drop_origin_fingerprints(struct blame_origin *o)
1006 if (o->fingerprints) {
1007 free_line_fingerprints(o->fingerprints, o->num_lines);
1008 o->num_lines = 0;
1009 FREE_AND_NULL(o->fingerprints);
1014 * Given an origin, prepare mmfile_t structure to be used by the
1015 * diff machinery
1017 static void fill_origin_blob(struct diff_options *opt,
1018 struct blame_origin *o, mmfile_t *file,
1019 int *num_read_blob, int fill_fingerprints)
1021 if (!o->file.ptr) {
1022 enum object_type type;
1023 unsigned long file_size;
1025 (*num_read_blob)++;
1026 if (opt->flags.allow_textconv &&
1027 textconv_object(opt->repo, o->path, o->mode,
1028 &o->blob_oid, 1, &file->ptr, &file_size))
1030 else
1031 file->ptr = read_object_file(&o->blob_oid, &type,
1032 &file_size);
1033 file->size = file_size;
1035 if (!file->ptr)
1036 die("Cannot read blob %s for path %s",
1037 oid_to_hex(&o->blob_oid),
1038 o->path);
1039 o->file = *file;
1041 else
1042 *file = o->file;
1043 if (fill_fingerprints)
1044 fill_origin_fingerprints(o);
1047 static void drop_origin_blob(struct blame_origin *o)
1049 FREE_AND_NULL(o->file.ptr);
1050 drop_origin_fingerprints(o);
1054 * Any merge of blames happens on lists of blames that arrived via
1055 * different parents in a single suspect. In this case, we want to
1056 * sort according to the suspect line numbers as opposed to the final
1057 * image line numbers. The function body is somewhat longish because
1058 * it avoids unnecessary writes.
1061 static struct blame_entry *blame_merge(struct blame_entry *list1,
1062 struct blame_entry *list2)
1064 struct blame_entry *p1 = list1, *p2 = list2,
1065 **tail = &list1;
1067 if (!p1)
1068 return p2;
1069 if (!p2)
1070 return p1;
1072 if (p1->s_lno <= p2->s_lno) {
1073 do {
1074 tail = &p1->next;
1075 if (!(p1 = *tail)) {
1076 *tail = p2;
1077 return list1;
1079 } while (p1->s_lno <= p2->s_lno);
1081 for (;;) {
1082 *tail = p2;
1083 do {
1084 tail = &p2->next;
1085 if (!(p2 = *tail)) {
1086 *tail = p1;
1087 return list1;
1089 } while (p1->s_lno > p2->s_lno);
1090 *tail = p1;
1091 do {
1092 tail = &p1->next;
1093 if (!(p1 = *tail)) {
1094 *tail = p2;
1095 return list1;
1097 } while (p1->s_lno <= p2->s_lno);
1101 static void *get_next_blame(const void *p)
1103 return ((struct blame_entry *)p)->next;
1106 static void set_next_blame(void *p1, void *p2)
1108 ((struct blame_entry *)p1)->next = p2;
1112 * Final image line numbers are all different, so we don't need a
1113 * three-way comparison here.
1116 static int compare_blame_final(const void *p1, const void *p2)
1118 return ((struct blame_entry *)p1)->lno > ((struct blame_entry *)p2)->lno
1119 ? 1 : -1;
1122 static int compare_blame_suspect(const void *p1, const void *p2)
1124 const struct blame_entry *s1 = p1, *s2 = p2;
1126 * to allow for collating suspects, we sort according to the
1127 * respective pointer value as the primary sorting criterion.
1128 * The actual relation is pretty unimportant as long as it
1129 * establishes a total order. Comparing as integers gives us
1130 * that.
1132 if (s1->suspect != s2->suspect)
1133 return (intptr_t)s1->suspect > (intptr_t)s2->suspect ? 1 : -1;
1134 if (s1->s_lno == s2->s_lno)
1135 return 0;
1136 return s1->s_lno > s2->s_lno ? 1 : -1;
1139 void blame_sort_final(struct blame_scoreboard *sb)
1141 sb->ent = llist_mergesort(sb->ent, get_next_blame, set_next_blame,
1142 compare_blame_final);
1145 static int compare_commits_by_reverse_commit_date(const void *a,
1146 const void *b,
1147 void *c)
1149 return -compare_commits_by_commit_date(a, b, c);
1153 * For debugging -- origin is refcounted, and this asserts that
1154 * we do not underflow.
1156 static void sanity_check_refcnt(struct blame_scoreboard *sb)
1158 int baa = 0;
1159 struct blame_entry *ent;
1161 for (ent = sb->ent; ent; ent = ent->next) {
1162 /* Nobody should have zero or negative refcnt */
1163 if (ent->suspect->refcnt <= 0) {
1164 fprintf(stderr, "%s in %s has negative refcnt %d\n",
1165 ent->suspect->path,
1166 oid_to_hex(&ent->suspect->commit->object.oid),
1167 ent->suspect->refcnt);
1168 baa = 1;
1171 if (baa)
1172 sb->on_sanity_fail(sb, baa);
1176 * If two blame entries that are next to each other came from
1177 * contiguous lines in the same origin (i.e. <commit, path> pair),
1178 * merge them together.
1180 void blame_coalesce(struct blame_scoreboard *sb)
1182 struct blame_entry *ent, *next;
1184 for (ent = sb->ent; ent && (next = ent->next); ent = next) {
1185 if (ent->suspect == next->suspect &&
1186 ent->s_lno + ent->num_lines == next->s_lno &&
1187 ent->lno + ent->num_lines == next->lno &&
1188 ent->ignored == next->ignored &&
1189 ent->unblamable == next->unblamable) {
1190 ent->num_lines += next->num_lines;
1191 ent->next = next->next;
1192 blame_origin_decref(next->suspect);
1193 free(next);
1194 ent->score = 0;
1195 next = ent; /* again */
1199 if (sb->debug) /* sanity */
1200 sanity_check_refcnt(sb);
1204 * Merge the given sorted list of blames into a preexisting origin.
1205 * If there were no previous blames to that commit, it is entered into
1206 * the commit priority queue of the score board.
1209 static void queue_blames(struct blame_scoreboard *sb, struct blame_origin *porigin,
1210 struct blame_entry *sorted)
1212 if (porigin->suspects)
1213 porigin->suspects = blame_merge(porigin->suspects, sorted);
1214 else {
1215 struct blame_origin *o;
1216 for (o = get_blame_suspects(porigin->commit); o; o = o->next) {
1217 if (o->suspects) {
1218 porigin->suspects = sorted;
1219 return;
1222 porigin->suspects = sorted;
1223 prio_queue_put(&sb->commits, porigin->commit);
1228 * Fill the blob_sha1 field of an origin if it hasn't, so that later
1229 * call to fill_origin_blob() can use it to locate the data. blob_sha1
1230 * for an origin is also used to pass the blame for the entire file to
1231 * the parent to detect the case where a child's blob is identical to
1232 * that of its parent's.
1234 * This also fills origin->mode for corresponding tree path.
1236 static int fill_blob_sha1_and_mode(struct repository *r,
1237 struct blame_origin *origin)
1239 if (!is_null_oid(&origin->blob_oid))
1240 return 0;
1241 if (get_tree_entry(r, &origin->commit->object.oid, origin->path, &origin->blob_oid, &origin->mode))
1242 goto error_out;
1243 if (oid_object_info(r, &origin->blob_oid, NULL) != OBJ_BLOB)
1244 goto error_out;
1245 return 0;
1246 error_out:
1247 oidclr(&origin->blob_oid);
1248 origin->mode = S_IFINVALID;
1249 return -1;
1252 struct blame_bloom_data {
1254 * Changed-path Bloom filter keys. These can help prevent
1255 * computing diffs against first parents, but we need to
1256 * expand the list as code is moved or files are renamed.
1258 struct bloom_filter_settings *settings;
1259 struct bloom_key **keys;
1260 int nr;
1261 int alloc;
1264 static int bloom_count_queries = 0;
1265 static int bloom_count_no = 0;
1266 static int maybe_changed_path(struct repository *r,
1267 struct blame_origin *origin,
1268 struct blame_bloom_data *bd)
1270 int i;
1271 struct bloom_filter *filter;
1273 if (!bd)
1274 return 1;
1276 if (commit_graph_generation(origin->commit) == GENERATION_NUMBER_INFINITY)
1277 return 1;
1279 filter = get_bloom_filter(r, origin->commit);
1281 if (!filter)
1282 return 1;
1284 bloom_count_queries++;
1285 for (i = 0; i < bd->nr; i++) {
1286 if (bloom_filter_contains(filter,
1287 bd->keys[i],
1288 bd->settings))
1289 return 1;
1292 bloom_count_no++;
1293 return 0;
1296 static void add_bloom_key(struct blame_bloom_data *bd,
1297 const char *path)
1299 if (!bd)
1300 return;
1302 if (bd->nr >= bd->alloc) {
1303 bd->alloc *= 2;
1304 REALLOC_ARRAY(bd->keys, bd->alloc);
1307 bd->keys[bd->nr] = xmalloc(sizeof(struct bloom_key));
1308 fill_bloom_key(path, strlen(path), bd->keys[bd->nr], bd->settings);
1309 bd->nr++;
1313 * We have an origin -- check if the same path exists in the
1314 * parent and return an origin structure to represent it.
1316 static struct blame_origin *find_origin(struct repository *r,
1317 struct commit *parent,
1318 struct blame_origin *origin,
1319 struct blame_bloom_data *bd)
1321 struct blame_origin *porigin;
1322 struct diff_options diff_opts;
1323 const char *paths[2];
1325 /* First check any existing origins */
1326 for (porigin = get_blame_suspects(parent); porigin; porigin = porigin->next)
1327 if (!strcmp(porigin->path, origin->path)) {
1329 * The same path between origin and its parent
1330 * without renaming -- the most common case.
1332 return blame_origin_incref (porigin);
1335 /* See if the origin->path is different between parent
1336 * and origin first. Most of the time they are the
1337 * same and diff-tree is fairly efficient about this.
1339 repo_diff_setup(r, &diff_opts);
1340 diff_opts.flags.recursive = 1;
1341 diff_opts.detect_rename = 0;
1342 diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1343 paths[0] = origin->path;
1344 paths[1] = NULL;
1346 parse_pathspec(&diff_opts.pathspec,
1347 PATHSPEC_ALL_MAGIC & ~PATHSPEC_LITERAL,
1348 PATHSPEC_LITERAL_PATH, "", paths);
1349 diff_setup_done(&diff_opts);
1351 if (is_null_oid(&origin->commit->object.oid))
1352 do_diff_cache(get_commit_tree_oid(parent), &diff_opts);
1353 else {
1354 int compute_diff = 1;
1355 if (origin->commit->parents &&
1356 oideq(&parent->object.oid,
1357 &origin->commit->parents->item->object.oid))
1358 compute_diff = maybe_changed_path(r, origin, bd);
1360 if (compute_diff)
1361 diff_tree_oid(get_commit_tree_oid(parent),
1362 get_commit_tree_oid(origin->commit),
1363 "", &diff_opts);
1365 diffcore_std(&diff_opts);
1367 if (!diff_queued_diff.nr) {
1368 /* The path is the same as parent */
1369 porigin = get_origin(parent, origin->path);
1370 oidcpy(&porigin->blob_oid, &origin->blob_oid);
1371 porigin->mode = origin->mode;
1372 } else {
1374 * Since origin->path is a pathspec, if the parent
1375 * commit had it as a directory, we will see a whole
1376 * bunch of deletion of files in the directory that we
1377 * do not care about.
1379 int i;
1380 struct diff_filepair *p = NULL;
1381 for (i = 0; i < diff_queued_diff.nr; i++) {
1382 const char *name;
1383 p = diff_queued_diff.queue[i];
1384 name = p->one->path ? p->one->path : p->two->path;
1385 if (!strcmp(name, origin->path))
1386 break;
1388 if (!p)
1389 die("internal error in blame::find_origin");
1390 switch (p->status) {
1391 default:
1392 die("internal error in blame::find_origin (%c)",
1393 p->status);
1394 case 'M':
1395 porigin = get_origin(parent, origin->path);
1396 oidcpy(&porigin->blob_oid, &p->one->oid);
1397 porigin->mode = p->one->mode;
1398 break;
1399 case 'A':
1400 case 'T':
1401 /* Did not exist in parent, or type changed */
1402 break;
1405 diff_flush(&diff_opts);
1406 return porigin;
1410 * We have an origin -- find the path that corresponds to it in its
1411 * parent and return an origin structure to represent it.
1413 static struct blame_origin *find_rename(struct repository *r,
1414 struct commit *parent,
1415 struct blame_origin *origin,
1416 struct blame_bloom_data *bd)
1418 struct blame_origin *porigin = NULL;
1419 struct diff_options diff_opts;
1420 int i;
1422 repo_diff_setup(r, &diff_opts);
1423 diff_opts.flags.recursive = 1;
1424 diff_opts.detect_rename = DIFF_DETECT_RENAME;
1425 diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1426 diff_opts.single_follow = origin->path;
1427 diff_setup_done(&diff_opts);
1429 if (is_null_oid(&origin->commit->object.oid))
1430 do_diff_cache(get_commit_tree_oid(parent), &diff_opts);
1431 else
1432 diff_tree_oid(get_commit_tree_oid(parent),
1433 get_commit_tree_oid(origin->commit),
1434 "", &diff_opts);
1435 diffcore_std(&diff_opts);
1437 for (i = 0; i < diff_queued_diff.nr; i++) {
1438 struct diff_filepair *p = diff_queued_diff.queue[i];
1439 if ((p->status == 'R' || p->status == 'C') &&
1440 !strcmp(p->two->path, origin->path)) {
1441 add_bloom_key(bd, p->one->path);
1442 porigin = get_origin(parent, p->one->path);
1443 oidcpy(&porigin->blob_oid, &p->one->oid);
1444 porigin->mode = p->one->mode;
1445 break;
1448 diff_flush(&diff_opts);
1449 return porigin;
1453 * Append a new blame entry to a given output queue.
1455 static void add_blame_entry(struct blame_entry ***queue,
1456 const struct blame_entry *src)
1458 struct blame_entry *e = xmalloc(sizeof(*e));
1459 memcpy(e, src, sizeof(*e));
1460 blame_origin_incref(e->suspect);
1462 e->next = **queue;
1463 **queue = e;
1464 *queue = &e->next;
1468 * src typically is on-stack; we want to copy the information in it to
1469 * a malloced blame_entry that gets added to the given queue. The
1470 * origin of dst loses a refcnt.
1472 static void dup_entry(struct blame_entry ***queue,
1473 struct blame_entry *dst, struct blame_entry *src)
1475 blame_origin_incref(src->suspect);
1476 blame_origin_decref(dst->suspect);
1477 memcpy(dst, src, sizeof(*src));
1478 dst->next = **queue;
1479 **queue = dst;
1480 *queue = &dst->next;
1483 const char *blame_nth_line(struct blame_scoreboard *sb, long lno)
1485 return sb->final_buf + sb->lineno[lno];
1489 * It is known that lines between tlno to same came from parent, and e
1490 * has an overlap with that range. it also is known that parent's
1491 * line plno corresponds to e's line tlno.
1493 * <---- e ----->
1494 * <------>
1495 * <------------>
1496 * <------------>
1497 * <------------------>
1499 * Split e into potentially three parts; before this chunk, the chunk
1500 * to be blamed for the parent, and after that portion.
1502 static void split_overlap(struct blame_entry *split,
1503 struct blame_entry *e,
1504 int tlno, int plno, int same,
1505 struct blame_origin *parent)
1507 int chunk_end_lno;
1508 int i;
1509 memset(split, 0, sizeof(struct blame_entry [3]));
1511 for (i = 0; i < 3; i++) {
1512 split[i].ignored = e->ignored;
1513 split[i].unblamable = e->unblamable;
1516 if (e->s_lno < tlno) {
1517 /* there is a pre-chunk part not blamed on parent */
1518 split[0].suspect = blame_origin_incref(e->suspect);
1519 split[0].lno = e->lno;
1520 split[0].s_lno = e->s_lno;
1521 split[0].num_lines = tlno - e->s_lno;
1522 split[1].lno = e->lno + tlno - e->s_lno;
1523 split[1].s_lno = plno;
1525 else {
1526 split[1].lno = e->lno;
1527 split[1].s_lno = plno + (e->s_lno - tlno);
1530 if (same < e->s_lno + e->num_lines) {
1531 /* there is a post-chunk part not blamed on parent */
1532 split[2].suspect = blame_origin_incref(e->suspect);
1533 split[2].lno = e->lno + (same - e->s_lno);
1534 split[2].s_lno = e->s_lno + (same - e->s_lno);
1535 split[2].num_lines = e->s_lno + e->num_lines - same;
1536 chunk_end_lno = split[2].lno;
1538 else
1539 chunk_end_lno = e->lno + e->num_lines;
1540 split[1].num_lines = chunk_end_lno - split[1].lno;
1543 * if it turns out there is nothing to blame the parent for,
1544 * forget about the splitting. !split[1].suspect signals this.
1546 if (split[1].num_lines < 1)
1547 return;
1548 split[1].suspect = blame_origin_incref(parent);
1552 * split_overlap() divided an existing blame e into up to three parts
1553 * in split. Any assigned blame is moved to queue to
1554 * reflect the split.
1556 static void split_blame(struct blame_entry ***blamed,
1557 struct blame_entry ***unblamed,
1558 struct blame_entry *split,
1559 struct blame_entry *e)
1561 if (split[0].suspect && split[2].suspect) {
1562 /* The first part (reuse storage for the existing entry e) */
1563 dup_entry(unblamed, e, &split[0]);
1565 /* The last part -- me */
1566 add_blame_entry(unblamed, &split[2]);
1568 /* ... and the middle part -- parent */
1569 add_blame_entry(blamed, &split[1]);
1571 else if (!split[0].suspect && !split[2].suspect)
1573 * The parent covers the entire area; reuse storage for
1574 * e and replace it with the parent.
1576 dup_entry(blamed, e, &split[1]);
1577 else if (split[0].suspect) {
1578 /* me and then parent */
1579 dup_entry(unblamed, e, &split[0]);
1580 add_blame_entry(blamed, &split[1]);
1582 else {
1583 /* parent and then me */
1584 dup_entry(blamed, e, &split[1]);
1585 add_blame_entry(unblamed, &split[2]);
1590 * After splitting the blame, the origins used by the
1591 * on-stack blame_entry should lose one refcnt each.
1593 static void decref_split(struct blame_entry *split)
1595 int i;
1597 for (i = 0; i < 3; i++)
1598 blame_origin_decref(split[i].suspect);
1602 * reverse_blame reverses the list given in head, appending tail.
1603 * That allows us to build lists in reverse order, then reverse them
1604 * afterwards. This can be faster than building the list in proper
1605 * order right away. The reason is that building in proper order
1606 * requires writing a link in the _previous_ element, while building
1607 * in reverse order just requires placing the list head into the
1608 * _current_ element.
1611 static struct blame_entry *reverse_blame(struct blame_entry *head,
1612 struct blame_entry *tail)
1614 while (head) {
1615 struct blame_entry *next = head->next;
1616 head->next = tail;
1617 tail = head;
1618 head = next;
1620 return tail;
1624 * Splits a blame entry into two entries at 'len' lines. The original 'e'
1625 * consists of len lines, i.e. [e->lno, e->lno + len), and the second part,
1626 * which is returned, consists of the remainder: [e->lno + len, e->lno +
1627 * e->num_lines). The caller needs to sort out the reference counting for the
1628 * new entry's suspect.
1630 static struct blame_entry *split_blame_at(struct blame_entry *e, int len,
1631 struct blame_origin *new_suspect)
1633 struct blame_entry *n = xcalloc(1, sizeof(struct blame_entry));
1635 n->suspect = new_suspect;
1636 n->ignored = e->ignored;
1637 n->unblamable = e->unblamable;
1638 n->lno = e->lno + len;
1639 n->s_lno = e->s_lno + len;
1640 n->num_lines = e->num_lines - len;
1641 e->num_lines = len;
1642 e->score = 0;
1643 return n;
1646 struct blame_line_tracker {
1647 int is_parent;
1648 int s_lno;
1651 static int are_lines_adjacent(struct blame_line_tracker *first,
1652 struct blame_line_tracker *second)
1654 return first->is_parent == second->is_parent &&
1655 first->s_lno + 1 == second->s_lno;
1658 static int scan_parent_range(struct fingerprint *p_fps,
1659 struct fingerprint *t_fps, int t_idx,
1660 int from, int nr_lines)
1662 int sim, p_idx;
1663 #define FINGERPRINT_FILE_THRESHOLD 10
1664 int best_sim_val = FINGERPRINT_FILE_THRESHOLD;
1665 int best_sim_idx = -1;
1667 for (p_idx = from; p_idx < from + nr_lines; p_idx++) {
1668 sim = fingerprint_similarity(&t_fps[t_idx], &p_fps[p_idx]);
1669 if (sim < best_sim_val)
1670 continue;
1671 /* Break ties with the closest-to-target line number */
1672 if (sim == best_sim_val && best_sim_idx != -1 &&
1673 abs(best_sim_idx - t_idx) < abs(p_idx - t_idx))
1674 continue;
1675 best_sim_val = sim;
1676 best_sim_idx = p_idx;
1678 return best_sim_idx;
1682 * The first pass checks the blame entry (from the target) against the parent's
1683 * diff chunk. If that fails for a line, the second pass tries to match that
1684 * line to any part of parent file. That catches cases where a change was
1685 * broken into two chunks by 'context.'
1687 static void guess_line_blames(struct blame_origin *parent,
1688 struct blame_origin *target,
1689 int tlno, int offset, int same, int parent_len,
1690 struct blame_line_tracker *line_blames)
1692 int i, best_idx, target_idx;
1693 int parent_slno = tlno + offset;
1694 int *fuzzy_matches;
1696 fuzzy_matches = fuzzy_find_matching_lines(parent, target,
1697 tlno, parent_slno, same,
1698 parent_len);
1699 for (i = 0; i < same - tlno; i++) {
1700 target_idx = tlno + i;
1701 if (fuzzy_matches && fuzzy_matches[i] >= 0) {
1702 best_idx = fuzzy_matches[i];
1703 } else {
1704 best_idx = scan_parent_range(parent->fingerprints,
1705 target->fingerprints,
1706 target_idx, 0,
1707 parent->num_lines);
1709 if (best_idx >= 0) {
1710 line_blames[i].is_parent = 1;
1711 line_blames[i].s_lno = best_idx;
1712 } else {
1713 line_blames[i].is_parent = 0;
1714 line_blames[i].s_lno = target_idx;
1717 free(fuzzy_matches);
1721 * This decides which parts of a blame entry go to the parent (added to the
1722 * ignoredp list) and which stay with the target (added to the diffp list). The
1723 * actual decision was made in a separate heuristic function, and those answers
1724 * for the lines in 'e' are in line_blames. This consumes e, essentially
1725 * putting it on a list.
1727 * Note that the blame entries on the ignoredp list are not necessarily sorted
1728 * with respect to the parent's line numbers yet.
1730 static void ignore_blame_entry(struct blame_entry *e,
1731 struct blame_origin *parent,
1732 struct blame_entry **diffp,
1733 struct blame_entry **ignoredp,
1734 struct blame_line_tracker *line_blames)
1736 int entry_len, nr_lines, i;
1739 * We carve new entries off the front of e. Each entry comes from a
1740 * contiguous chunk of lines: adjacent lines from the same origin
1741 * (either the parent or the target).
1743 entry_len = 1;
1744 nr_lines = e->num_lines; /* e changes in the loop */
1745 for (i = 0; i < nr_lines; i++) {
1746 struct blame_entry *next = NULL;
1749 * We are often adjacent to the next line - only split the blame
1750 * entry when we have to.
1752 if (i + 1 < nr_lines) {
1753 if (are_lines_adjacent(&line_blames[i],
1754 &line_blames[i + 1])) {
1755 entry_len++;
1756 continue;
1758 next = split_blame_at(e, entry_len,
1759 blame_origin_incref(e->suspect));
1761 if (line_blames[i].is_parent) {
1762 e->ignored = 1;
1763 blame_origin_decref(e->suspect);
1764 e->suspect = blame_origin_incref(parent);
1765 e->s_lno = line_blames[i - entry_len + 1].s_lno;
1766 e->next = *ignoredp;
1767 *ignoredp = e;
1768 } else {
1769 e->unblamable = 1;
1770 /* e->s_lno is already in the target's address space. */
1771 e->next = *diffp;
1772 *diffp = e;
1774 assert(e->num_lines == entry_len);
1775 e = next;
1776 entry_len = 1;
1778 assert(!e);
1782 * Process one hunk from the patch between the current suspect for
1783 * blame_entry e and its parent. This first blames any unfinished
1784 * entries before the chunk (which is where target and parent start
1785 * differing) on the parent, and then splits blame entries at the
1786 * start and at the end of the difference region. Since use of -M and
1787 * -C options may lead to overlapping/duplicate source line number
1788 * ranges, all we can rely on from sorting/merging is the order of the
1789 * first suspect line number.
1791 * tlno: line number in the target where this chunk begins
1792 * same: line number in the target where this chunk ends
1793 * offset: add to tlno to get the chunk starting point in the parent
1794 * parent_len: number of lines in the parent chunk
1796 static void blame_chunk(struct blame_entry ***dstq, struct blame_entry ***srcq,
1797 int tlno, int offset, int same, int parent_len,
1798 struct blame_origin *parent,
1799 struct blame_origin *target, int ignore_diffs)
1801 struct blame_entry *e = **srcq;
1802 struct blame_entry *samep = NULL, *diffp = NULL, *ignoredp = NULL;
1803 struct blame_line_tracker *line_blames = NULL;
1805 while (e && e->s_lno < tlno) {
1806 struct blame_entry *next = e->next;
1808 * current record starts before differing portion. If
1809 * it reaches into it, we need to split it up and
1810 * examine the second part separately.
1812 if (e->s_lno + e->num_lines > tlno) {
1813 /* Move second half to a new record */
1814 struct blame_entry *n;
1816 n = split_blame_at(e, tlno - e->s_lno, e->suspect);
1817 /* Push new record to diffp */
1818 n->next = diffp;
1819 diffp = n;
1820 } else
1821 blame_origin_decref(e->suspect);
1822 /* Pass blame for everything before the differing
1823 * chunk to the parent */
1824 e->suspect = blame_origin_incref(parent);
1825 e->s_lno += offset;
1826 e->next = samep;
1827 samep = e;
1828 e = next;
1831 * As we don't know how much of a common stretch after this
1832 * diff will occur, the currently blamed parts are all that we
1833 * can assign to the parent for now.
1836 if (samep) {
1837 **dstq = reverse_blame(samep, **dstq);
1838 *dstq = &samep->next;
1841 * Prepend the split off portions: everything after e starts
1842 * after the blameable portion.
1844 e = reverse_blame(diffp, e);
1847 * Now retain records on the target while parts are different
1848 * from the parent.
1850 samep = NULL;
1851 diffp = NULL;
1853 if (ignore_diffs && same - tlno > 0) {
1854 CALLOC_ARRAY(line_blames, same - tlno);
1855 guess_line_blames(parent, target, tlno, offset, same,
1856 parent_len, line_blames);
1859 while (e && e->s_lno < same) {
1860 struct blame_entry *next = e->next;
1863 * If current record extends into sameness, need to split.
1865 if (e->s_lno + e->num_lines > same) {
1867 * Move second half to a new record to be
1868 * processed by later chunks
1870 struct blame_entry *n;
1872 n = split_blame_at(e, same - e->s_lno,
1873 blame_origin_incref(e->suspect));
1874 /* Push new record to samep */
1875 n->next = samep;
1876 samep = n;
1878 if (ignore_diffs) {
1879 ignore_blame_entry(e, parent, &diffp, &ignoredp,
1880 line_blames + e->s_lno - tlno);
1881 } else {
1882 e->next = diffp;
1883 diffp = e;
1885 e = next;
1887 free(line_blames);
1888 if (ignoredp) {
1890 * Note ignoredp is not sorted yet, and thus neither is dstq.
1891 * That list must be sorted before we queue_blames(). We defer
1892 * sorting until after all diff hunks are processed, so that
1893 * guess_line_blames() can pick *any* line in the parent. The
1894 * slight drawback is that we end up sorting all blame entries
1895 * passed to the parent, including those that are unrelated to
1896 * changes made by the ignored commit.
1898 **dstq = reverse_blame(ignoredp, **dstq);
1899 *dstq = &ignoredp->next;
1901 **srcq = reverse_blame(diffp, reverse_blame(samep, e));
1902 /* Move across elements that are in the unblamable portion */
1903 if (diffp)
1904 *srcq = &diffp->next;
1907 struct blame_chunk_cb_data {
1908 struct blame_origin *parent;
1909 struct blame_origin *target;
1910 long offset;
1911 int ignore_diffs;
1912 struct blame_entry **dstq;
1913 struct blame_entry **srcq;
1916 /* diff chunks are from parent to target */
1917 static int blame_chunk_cb(long start_a, long count_a,
1918 long start_b, long count_b, void *data)
1920 struct blame_chunk_cb_data *d = data;
1921 if (start_a - start_b != d->offset)
1922 die("internal error in blame::blame_chunk_cb");
1923 blame_chunk(&d->dstq, &d->srcq, start_b, start_a - start_b,
1924 start_b + count_b, count_a, d->parent, d->target,
1925 d->ignore_diffs);
1926 d->offset = start_a + count_a - (start_b + count_b);
1927 return 0;
1931 * We are looking at the origin 'target' and aiming to pass blame
1932 * for the lines it is suspected to its parent. Run diff to find
1933 * which lines came from parent and pass blame for them.
1935 static void pass_blame_to_parent(struct blame_scoreboard *sb,
1936 struct blame_origin *target,
1937 struct blame_origin *parent, int ignore_diffs)
1939 mmfile_t file_p, file_o;
1940 struct blame_chunk_cb_data d;
1941 struct blame_entry *newdest = NULL;
1943 if (!target->suspects)
1944 return; /* nothing remains for this target */
1946 d.parent = parent;
1947 d.target = target;
1948 d.offset = 0;
1949 d.ignore_diffs = ignore_diffs;
1950 d.dstq = &newdest; d.srcq = &target->suspects;
1952 fill_origin_blob(&sb->revs->diffopt, parent, &file_p,
1953 &sb->num_read_blob, ignore_diffs);
1954 fill_origin_blob(&sb->revs->diffopt, target, &file_o,
1955 &sb->num_read_blob, ignore_diffs);
1956 sb->num_get_patch++;
1958 if (diff_hunks(&file_p, &file_o, blame_chunk_cb, &d, sb->xdl_opts))
1959 die("unable to generate diff (%s -> %s)",
1960 oid_to_hex(&parent->commit->object.oid),
1961 oid_to_hex(&target->commit->object.oid));
1962 /* The rest are the same as the parent */
1963 blame_chunk(&d.dstq, &d.srcq, INT_MAX, d.offset, INT_MAX, 0,
1964 parent, target, 0);
1965 *d.dstq = NULL;
1966 if (ignore_diffs)
1967 newdest = llist_mergesort(newdest, get_next_blame,
1968 set_next_blame,
1969 compare_blame_suspect);
1970 queue_blames(sb, parent, newdest);
1972 return;
1976 * The lines in blame_entry after splitting blames many times can become
1977 * very small and trivial, and at some point it becomes pointless to
1978 * blame the parents. E.g. "\t\t}\n\t}\n\n" appears everywhere in any
1979 * ordinary C program, and it is not worth to say it was copied from
1980 * totally unrelated file in the parent.
1982 * Compute how trivial the lines in the blame_entry are.
1984 unsigned blame_entry_score(struct blame_scoreboard *sb, struct blame_entry *e)
1986 unsigned score;
1987 const char *cp, *ep;
1989 if (e->score)
1990 return e->score;
1992 score = 1;
1993 cp = blame_nth_line(sb, e->lno);
1994 ep = blame_nth_line(sb, e->lno + e->num_lines);
1995 while (cp < ep) {
1996 unsigned ch = *((unsigned char *)cp);
1997 if (isalnum(ch))
1998 score++;
1999 cp++;
2001 e->score = score;
2002 return score;
2006 * best_so_far[] and potential[] are both a split of an existing blame_entry
2007 * that passes blame to the parent. Maintain best_so_far the best split so
2008 * far, by comparing potential and best_so_far and copying potential into
2009 * bst_so_far as needed.
2011 static void copy_split_if_better(struct blame_scoreboard *sb,
2012 struct blame_entry *best_so_far,
2013 struct blame_entry *potential)
2015 int i;
2017 if (!potential[1].suspect)
2018 return;
2019 if (best_so_far[1].suspect) {
2020 if (blame_entry_score(sb, &potential[1]) <
2021 blame_entry_score(sb, &best_so_far[1]))
2022 return;
2025 for (i = 0; i < 3; i++)
2026 blame_origin_incref(potential[i].suspect);
2027 decref_split(best_so_far);
2028 memcpy(best_so_far, potential, sizeof(struct blame_entry[3]));
2032 * We are looking at a part of the final image represented by
2033 * ent (tlno and same are offset by ent->s_lno).
2034 * tlno is where we are looking at in the final image.
2035 * up to (but not including) same match preimage.
2036 * plno is where we are looking at in the preimage.
2038 * <-------------- final image ---------------------->
2039 * <------ent------>
2040 * ^tlno ^same
2041 * <---------preimage----->
2042 * ^plno
2044 * All line numbers are 0-based.
2046 static void handle_split(struct blame_scoreboard *sb,
2047 struct blame_entry *ent,
2048 int tlno, int plno, int same,
2049 struct blame_origin *parent,
2050 struct blame_entry *split)
2052 if (ent->num_lines <= tlno)
2053 return;
2054 if (tlno < same) {
2055 struct blame_entry potential[3];
2056 tlno += ent->s_lno;
2057 same += ent->s_lno;
2058 split_overlap(potential, ent, tlno, plno, same, parent);
2059 copy_split_if_better(sb, split, potential);
2060 decref_split(potential);
2064 struct handle_split_cb_data {
2065 struct blame_scoreboard *sb;
2066 struct blame_entry *ent;
2067 struct blame_origin *parent;
2068 struct blame_entry *split;
2069 long plno;
2070 long tlno;
2073 static int handle_split_cb(long start_a, long count_a,
2074 long start_b, long count_b, void *data)
2076 struct handle_split_cb_data *d = data;
2077 handle_split(d->sb, d->ent, d->tlno, d->plno, start_b, d->parent,
2078 d->split);
2079 d->plno = start_a + count_a;
2080 d->tlno = start_b + count_b;
2081 return 0;
2085 * Find the lines from parent that are the same as ent so that
2086 * we can pass blames to it. file_p has the blob contents for
2087 * the parent.
2089 static void find_copy_in_blob(struct blame_scoreboard *sb,
2090 struct blame_entry *ent,
2091 struct blame_origin *parent,
2092 struct blame_entry *split,
2093 mmfile_t *file_p)
2095 const char *cp;
2096 mmfile_t file_o;
2097 struct handle_split_cb_data d;
2099 memset(&d, 0, sizeof(d));
2100 d.sb = sb; d.ent = ent; d.parent = parent; d.split = split;
2102 * Prepare mmfile that contains only the lines in ent.
2104 cp = blame_nth_line(sb, ent->lno);
2105 file_o.ptr = (char *) cp;
2106 file_o.size = blame_nth_line(sb, ent->lno + ent->num_lines) - cp;
2109 * file_o is a part of final image we are annotating.
2110 * file_p partially may match that image.
2112 memset(split, 0, sizeof(struct blame_entry [3]));
2113 if (diff_hunks(file_p, &file_o, handle_split_cb, &d, sb->xdl_opts))
2114 die("unable to generate diff (%s)",
2115 oid_to_hex(&parent->commit->object.oid));
2116 /* remainder, if any, all match the preimage */
2117 handle_split(sb, ent, d.tlno, d.plno, ent->num_lines, parent, split);
2120 /* Move all blame entries from list *source that have a score smaller
2121 * than score_min to the front of list *small.
2122 * Returns a pointer to the link pointing to the old head of the small list.
2125 static struct blame_entry **filter_small(struct blame_scoreboard *sb,
2126 struct blame_entry **small,
2127 struct blame_entry **source,
2128 unsigned score_min)
2130 struct blame_entry *p = *source;
2131 struct blame_entry *oldsmall = *small;
2132 while (p) {
2133 if (blame_entry_score(sb, p) <= score_min) {
2134 *small = p;
2135 small = &p->next;
2136 p = *small;
2137 } else {
2138 *source = p;
2139 source = &p->next;
2140 p = *source;
2143 *small = oldsmall;
2144 *source = NULL;
2145 return small;
2149 * See if lines currently target is suspected for can be attributed to
2150 * parent.
2152 static void find_move_in_parent(struct blame_scoreboard *sb,
2153 struct blame_entry ***blamed,
2154 struct blame_entry **toosmall,
2155 struct blame_origin *target,
2156 struct blame_origin *parent)
2158 struct blame_entry *e, split[3];
2159 struct blame_entry *unblamed = target->suspects;
2160 struct blame_entry *leftover = NULL;
2161 mmfile_t file_p;
2163 if (!unblamed)
2164 return; /* nothing remains for this target */
2166 fill_origin_blob(&sb->revs->diffopt, parent, &file_p,
2167 &sb->num_read_blob, 0);
2168 if (!file_p.ptr)
2169 return;
2171 /* At each iteration, unblamed has a NULL-terminated list of
2172 * entries that have not yet been tested for blame. leftover
2173 * contains the reversed list of entries that have been tested
2174 * without being assignable to the parent.
2176 do {
2177 struct blame_entry **unblamedtail = &unblamed;
2178 struct blame_entry *next;
2179 for (e = unblamed; e; e = next) {
2180 next = e->next;
2181 find_copy_in_blob(sb, e, parent, split, &file_p);
2182 if (split[1].suspect &&
2183 sb->move_score < blame_entry_score(sb, &split[1])) {
2184 split_blame(blamed, &unblamedtail, split, e);
2185 } else {
2186 e->next = leftover;
2187 leftover = e;
2189 decref_split(split);
2191 *unblamedtail = NULL;
2192 toosmall = filter_small(sb, toosmall, &unblamed, sb->move_score);
2193 } while (unblamed);
2194 target->suspects = reverse_blame(leftover, NULL);
2197 struct blame_list {
2198 struct blame_entry *ent;
2199 struct blame_entry split[3];
2203 * Count the number of entries the target is suspected for,
2204 * and prepare a list of entry and the best split.
2206 static struct blame_list *setup_blame_list(struct blame_entry *unblamed,
2207 int *num_ents_p)
2209 struct blame_entry *e;
2210 int num_ents, i;
2211 struct blame_list *blame_list = NULL;
2213 for (e = unblamed, num_ents = 0; e; e = e->next)
2214 num_ents++;
2215 if (num_ents) {
2216 CALLOC_ARRAY(blame_list, num_ents);
2217 for (e = unblamed, i = 0; e; e = e->next)
2218 blame_list[i++].ent = e;
2220 *num_ents_p = num_ents;
2221 return blame_list;
2225 * For lines target is suspected for, see if we can find code movement
2226 * across file boundary from the parent commit. porigin is the path
2227 * in the parent we already tried.
2229 static void find_copy_in_parent(struct blame_scoreboard *sb,
2230 struct blame_entry ***blamed,
2231 struct blame_entry **toosmall,
2232 struct blame_origin *target,
2233 struct commit *parent,
2234 struct blame_origin *porigin,
2235 int opt)
2237 struct diff_options diff_opts;
2238 int i, j;
2239 struct blame_list *blame_list;
2240 int num_ents;
2241 struct blame_entry *unblamed = target->suspects;
2242 struct blame_entry *leftover = NULL;
2244 if (!unblamed)
2245 return; /* nothing remains for this target */
2247 repo_diff_setup(sb->repo, &diff_opts);
2248 diff_opts.flags.recursive = 1;
2249 diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
2251 diff_setup_done(&diff_opts);
2253 /* Try "find copies harder" on new path if requested;
2254 * we do not want to use diffcore_rename() actually to
2255 * match things up; find_copies_harder is set only to
2256 * force diff_tree_oid() to feed all filepairs to diff_queue,
2257 * and this code needs to be after diff_setup_done(), which
2258 * usually makes find-copies-harder imply copy detection.
2260 if ((opt & PICKAXE_BLAME_COPY_HARDEST)
2261 || ((opt & PICKAXE_BLAME_COPY_HARDER)
2262 && (!porigin || strcmp(target->path, porigin->path))))
2263 diff_opts.flags.find_copies_harder = 1;
2265 if (is_null_oid(&target->commit->object.oid))
2266 do_diff_cache(get_commit_tree_oid(parent), &diff_opts);
2267 else
2268 diff_tree_oid(get_commit_tree_oid(parent),
2269 get_commit_tree_oid(target->commit),
2270 "", &diff_opts);
2272 if (!diff_opts.flags.find_copies_harder)
2273 diffcore_std(&diff_opts);
2275 do {
2276 struct blame_entry **unblamedtail = &unblamed;
2277 blame_list = setup_blame_list(unblamed, &num_ents);
2279 for (i = 0; i < diff_queued_diff.nr; i++) {
2280 struct diff_filepair *p = diff_queued_diff.queue[i];
2281 struct blame_origin *norigin;
2282 mmfile_t file_p;
2283 struct blame_entry potential[3];
2285 if (!DIFF_FILE_VALID(p->one))
2286 continue; /* does not exist in parent */
2287 if (S_ISGITLINK(p->one->mode))
2288 continue; /* ignore git links */
2289 if (porigin && !strcmp(p->one->path, porigin->path))
2290 /* find_move already dealt with this path */
2291 continue;
2293 norigin = get_origin(parent, p->one->path);
2294 oidcpy(&norigin->blob_oid, &p->one->oid);
2295 norigin->mode = p->one->mode;
2296 fill_origin_blob(&sb->revs->diffopt, norigin, &file_p,
2297 &sb->num_read_blob, 0);
2298 if (!file_p.ptr)
2299 continue;
2301 for (j = 0; j < num_ents; j++) {
2302 find_copy_in_blob(sb, blame_list[j].ent,
2303 norigin, potential, &file_p);
2304 copy_split_if_better(sb, blame_list[j].split,
2305 potential);
2306 decref_split(potential);
2308 blame_origin_decref(norigin);
2311 for (j = 0; j < num_ents; j++) {
2312 struct blame_entry *split = blame_list[j].split;
2313 if (split[1].suspect &&
2314 sb->copy_score < blame_entry_score(sb, &split[1])) {
2315 split_blame(blamed, &unblamedtail, split,
2316 blame_list[j].ent);
2317 } else {
2318 blame_list[j].ent->next = leftover;
2319 leftover = blame_list[j].ent;
2321 decref_split(split);
2323 free(blame_list);
2324 *unblamedtail = NULL;
2325 toosmall = filter_small(sb, toosmall, &unblamed, sb->copy_score);
2326 } while (unblamed);
2327 target->suspects = reverse_blame(leftover, NULL);
2328 diff_flush(&diff_opts);
2332 * The blobs of origin and porigin exactly match, so everything
2333 * origin is suspected for can be blamed on the parent.
2335 static void pass_whole_blame(struct blame_scoreboard *sb,
2336 struct blame_origin *origin, struct blame_origin *porigin)
2338 struct blame_entry *e, *suspects;
2340 if (!porigin->file.ptr && origin->file.ptr) {
2341 /* Steal its file */
2342 porigin->file = origin->file;
2343 origin->file.ptr = NULL;
2345 suspects = origin->suspects;
2346 origin->suspects = NULL;
2347 for (e = suspects; e; e = e->next) {
2348 blame_origin_incref(porigin);
2349 blame_origin_decref(e->suspect);
2350 e->suspect = porigin;
2352 queue_blames(sb, porigin, suspects);
2356 * We pass blame from the current commit to its parents. We keep saying
2357 * "parent" (and "porigin"), but what we mean is to find scapegoat to
2358 * exonerate ourselves.
2360 static struct commit_list *first_scapegoat(struct rev_info *revs, struct commit *commit,
2361 int reverse)
2363 if (!reverse) {
2364 if (revs->first_parent_only &&
2365 commit->parents &&
2366 commit->parents->next) {
2367 free_commit_list(commit->parents->next);
2368 commit->parents->next = NULL;
2370 return commit->parents;
2372 return lookup_decoration(&revs->children, &commit->object);
2375 static int num_scapegoats(struct rev_info *revs, struct commit *commit, int reverse)
2377 struct commit_list *l = first_scapegoat(revs, commit, reverse);
2378 return commit_list_count(l);
2381 /* Distribute collected unsorted blames to the respected sorted lists
2382 * in the various origins.
2384 static void distribute_blame(struct blame_scoreboard *sb, struct blame_entry *blamed)
2386 blamed = llist_mergesort(blamed, get_next_blame, set_next_blame,
2387 compare_blame_suspect);
2388 while (blamed)
2390 struct blame_origin *porigin = blamed->suspect;
2391 struct blame_entry *suspects = NULL;
2392 do {
2393 struct blame_entry *next = blamed->next;
2394 blamed->next = suspects;
2395 suspects = blamed;
2396 blamed = next;
2397 } while (blamed && blamed->suspect == porigin);
2398 suspects = reverse_blame(suspects, NULL);
2399 queue_blames(sb, porigin, suspects);
2403 #define MAXSG 16
2405 typedef struct blame_origin *(*blame_find_alg)(struct repository *,
2406 struct commit *,
2407 struct blame_origin *,
2408 struct blame_bloom_data *);
2410 static void pass_blame(struct blame_scoreboard *sb, struct blame_origin *origin, int opt)
2412 struct rev_info *revs = sb->revs;
2413 int i, pass, num_sg;
2414 struct commit *commit = origin->commit;
2415 struct commit_list *sg;
2416 struct blame_origin *sg_buf[MAXSG];
2417 struct blame_origin *porigin, **sg_origin = sg_buf;
2418 struct blame_entry *toosmall = NULL;
2419 struct blame_entry *blames, **blametail = &blames;
2421 num_sg = num_scapegoats(revs, commit, sb->reverse);
2422 if (!num_sg)
2423 goto finish;
2424 else if (num_sg < ARRAY_SIZE(sg_buf))
2425 memset(sg_buf, 0, sizeof(sg_buf));
2426 else
2427 CALLOC_ARRAY(sg_origin, num_sg);
2430 * The first pass looks for unrenamed path to optimize for
2431 * common cases, then we look for renames in the second pass.
2433 for (pass = 0; pass < 2 - sb->no_whole_file_rename; pass++) {
2434 blame_find_alg find = pass ? find_rename : find_origin;
2436 for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
2437 i < num_sg && sg;
2438 sg = sg->next, i++) {
2439 struct commit *p = sg->item;
2440 int j, same;
2442 if (sg_origin[i])
2443 continue;
2444 if (parse_commit(p))
2445 continue;
2446 porigin = find(sb->repo, p, origin, sb->bloom_data);
2447 if (!porigin)
2448 continue;
2449 if (oideq(&porigin->blob_oid, &origin->blob_oid)) {
2450 pass_whole_blame(sb, origin, porigin);
2451 blame_origin_decref(porigin);
2452 goto finish;
2454 for (j = same = 0; j < i; j++)
2455 if (sg_origin[j] &&
2456 oideq(&sg_origin[j]->blob_oid, &porigin->blob_oid)) {
2457 same = 1;
2458 break;
2460 if (!same)
2461 sg_origin[i] = porigin;
2462 else
2463 blame_origin_decref(porigin);
2467 sb->num_commits++;
2468 for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
2469 i < num_sg && sg;
2470 sg = sg->next, i++) {
2471 struct blame_origin *porigin = sg_origin[i];
2472 if (!porigin)
2473 continue;
2474 if (!origin->previous) {
2475 blame_origin_incref(porigin);
2476 origin->previous = porigin;
2478 pass_blame_to_parent(sb, origin, porigin, 0);
2479 if (!origin->suspects)
2480 goto finish;
2484 * Pass remaining suspects for ignored commits to their parents.
2486 if (oidset_contains(&sb->ignore_list, &commit->object.oid)) {
2487 for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
2488 i < num_sg && sg;
2489 sg = sg->next, i++) {
2490 struct blame_origin *porigin = sg_origin[i];
2492 if (!porigin)
2493 continue;
2494 pass_blame_to_parent(sb, origin, porigin, 1);
2496 * Preemptively drop porigin so we can refresh the
2497 * fingerprints if we use the parent again, which can
2498 * occur if you ignore back-to-back commits.
2500 drop_origin_blob(porigin);
2501 if (!origin->suspects)
2502 goto finish;
2507 * Optionally find moves in parents' files.
2509 if (opt & PICKAXE_BLAME_MOVE) {
2510 filter_small(sb, &toosmall, &origin->suspects, sb->move_score);
2511 if (origin->suspects) {
2512 for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
2513 i < num_sg && sg;
2514 sg = sg->next, i++) {
2515 struct blame_origin *porigin = sg_origin[i];
2516 if (!porigin)
2517 continue;
2518 find_move_in_parent(sb, &blametail, &toosmall, origin, porigin);
2519 if (!origin->suspects)
2520 break;
2526 * Optionally find copies from parents' files.
2528 if (opt & PICKAXE_BLAME_COPY) {
2529 if (sb->copy_score > sb->move_score)
2530 filter_small(sb, &toosmall, &origin->suspects, sb->copy_score);
2531 else if (sb->copy_score < sb->move_score) {
2532 origin->suspects = blame_merge(origin->suspects, toosmall);
2533 toosmall = NULL;
2534 filter_small(sb, &toosmall, &origin->suspects, sb->copy_score);
2536 if (!origin->suspects)
2537 goto finish;
2539 for (i = 0, sg = first_scapegoat(revs, commit, sb->reverse);
2540 i < num_sg && sg;
2541 sg = sg->next, i++) {
2542 struct blame_origin *porigin = sg_origin[i];
2543 find_copy_in_parent(sb, &blametail, &toosmall,
2544 origin, sg->item, porigin, opt);
2545 if (!origin->suspects)
2546 goto finish;
2550 finish:
2551 *blametail = NULL;
2552 distribute_blame(sb, blames);
2554 * prepend toosmall to origin->suspects
2556 * There is no point in sorting: this ends up on a big
2557 * unsorted list in the caller anyway.
2559 if (toosmall) {
2560 struct blame_entry **tail = &toosmall;
2561 while (*tail)
2562 tail = &(*tail)->next;
2563 *tail = origin->suspects;
2564 origin->suspects = toosmall;
2566 for (i = 0; i < num_sg; i++) {
2567 if (sg_origin[i]) {
2568 if (!sg_origin[i]->suspects)
2569 drop_origin_blob(sg_origin[i]);
2570 blame_origin_decref(sg_origin[i]);
2573 drop_origin_blob(origin);
2574 if (sg_buf != sg_origin)
2575 free(sg_origin);
2579 * The main loop -- while we have blobs with lines whose true origin
2580 * is still unknown, pick one blob, and allow its lines to pass blames
2581 * to its parents. */
2582 void assign_blame(struct blame_scoreboard *sb, int opt)
2584 struct rev_info *revs = sb->revs;
2585 struct commit *commit = prio_queue_get(&sb->commits);
2587 while (commit) {
2588 struct blame_entry *ent;
2589 struct blame_origin *suspect = get_blame_suspects(commit);
2591 /* find one suspect to break down */
2592 while (suspect && !suspect->suspects)
2593 suspect = suspect->next;
2595 if (!suspect) {
2596 commit = prio_queue_get(&sb->commits);
2597 continue;
2600 assert(commit == suspect->commit);
2603 * We will use this suspect later in the loop,
2604 * so hold onto it in the meantime.
2606 blame_origin_incref(suspect);
2607 parse_commit(commit);
2608 if (sb->reverse ||
2609 (!(commit->object.flags & UNINTERESTING) &&
2610 !(revs->max_age != -1 && commit->date < revs->max_age)))
2611 pass_blame(sb, suspect, opt);
2612 else {
2613 commit->object.flags |= UNINTERESTING;
2614 if (commit->object.parsed)
2615 mark_parents_uninteresting(sb->revs, commit);
2617 /* treat root commit as boundary */
2618 if (!commit->parents && !sb->show_root)
2619 commit->object.flags |= UNINTERESTING;
2621 /* Take responsibility for the remaining entries */
2622 ent = suspect->suspects;
2623 if (ent) {
2624 suspect->guilty = 1;
2625 for (;;) {
2626 struct blame_entry *next = ent->next;
2627 if (sb->found_guilty_entry)
2628 sb->found_guilty_entry(ent, sb->found_guilty_entry_data);
2629 if (next) {
2630 ent = next;
2631 continue;
2633 ent->next = sb->ent;
2634 sb->ent = suspect->suspects;
2635 suspect->suspects = NULL;
2636 break;
2639 blame_origin_decref(suspect);
2641 if (sb->debug) /* sanity */
2642 sanity_check_refcnt(sb);
2647 * To allow quick access to the contents of nth line in the
2648 * final image, prepare an index in the scoreboard.
2650 static int prepare_lines(struct blame_scoreboard *sb)
2652 sb->num_lines = find_line_starts(&sb->lineno, sb->final_buf,
2653 sb->final_buf_size);
2654 return sb->num_lines;
2657 static struct commit *find_single_final(struct rev_info *revs,
2658 const char **name_p)
2660 int i;
2661 struct commit *found = NULL;
2662 const char *name = NULL;
2664 for (i = 0; i < revs->pending.nr; i++) {
2665 struct object *obj = revs->pending.objects[i].item;
2666 if (obj->flags & UNINTERESTING)
2667 continue;
2668 obj = deref_tag(revs->repo, obj, NULL, 0);
2669 if (!obj || obj->type != OBJ_COMMIT)
2670 die("Non commit %s?", revs->pending.objects[i].name);
2671 if (found)
2672 die("More than one commit to dig from %s and %s?",
2673 revs->pending.objects[i].name, name);
2674 found = (struct commit *)obj;
2675 name = revs->pending.objects[i].name;
2677 if (name_p)
2678 *name_p = xstrdup_or_null(name);
2679 return found;
2682 static struct commit *dwim_reverse_initial(struct rev_info *revs,
2683 const char **name_p)
2686 * DWIM "git blame --reverse ONE -- PATH" as
2687 * "git blame --reverse ONE..HEAD -- PATH" but only do so
2688 * when it makes sense.
2690 struct object *obj;
2691 struct commit *head_commit;
2692 struct object_id head_oid;
2694 if (revs->pending.nr != 1)
2695 return NULL;
2697 /* Is that sole rev a committish? */
2698 obj = revs->pending.objects[0].item;
2699 obj = deref_tag(revs->repo, obj, NULL, 0);
2700 if (!obj || obj->type != OBJ_COMMIT)
2701 return NULL;
2703 /* Do we have HEAD? */
2704 if (!resolve_ref_unsafe("HEAD", RESOLVE_REF_READING, &head_oid, NULL))
2705 return NULL;
2706 head_commit = lookup_commit_reference_gently(revs->repo,
2707 &head_oid, 1);
2708 if (!head_commit)
2709 return NULL;
2711 /* Turn "ONE" into "ONE..HEAD" then */
2712 obj->flags |= UNINTERESTING;
2713 add_pending_object(revs, &head_commit->object, "HEAD");
2715 if (name_p)
2716 *name_p = revs->pending.objects[0].name;
2717 return (struct commit *)obj;
2720 static struct commit *find_single_initial(struct rev_info *revs,
2721 const char **name_p)
2723 int i;
2724 struct commit *found = NULL;
2725 const char *name = NULL;
2728 * There must be one and only one negative commit, and it must be
2729 * the boundary.
2731 for (i = 0; i < revs->pending.nr; i++) {
2732 struct object *obj = revs->pending.objects[i].item;
2733 if (!(obj->flags & UNINTERESTING))
2734 continue;
2735 obj = deref_tag(revs->repo, obj, NULL, 0);
2736 if (!obj || obj->type != OBJ_COMMIT)
2737 die("Non commit %s?", revs->pending.objects[i].name);
2738 if (found)
2739 die("More than one commit to dig up from, %s and %s?",
2740 revs->pending.objects[i].name, name);
2741 found = (struct commit *) obj;
2742 name = revs->pending.objects[i].name;
2745 if (!name)
2746 found = dwim_reverse_initial(revs, &name);
2747 if (!name)
2748 die("No commit to dig up from?");
2750 if (name_p)
2751 *name_p = xstrdup(name);
2752 return found;
2755 void init_scoreboard(struct blame_scoreboard *sb)
2757 memset(sb, 0, sizeof(struct blame_scoreboard));
2758 sb->move_score = BLAME_DEFAULT_MOVE_SCORE;
2759 sb->copy_score = BLAME_DEFAULT_COPY_SCORE;
2762 void setup_scoreboard(struct blame_scoreboard *sb,
2763 struct blame_origin **orig)
2765 const char *final_commit_name = NULL;
2766 struct blame_origin *o;
2767 struct commit *final_commit = NULL;
2768 enum object_type type;
2770 init_blame_suspects(&blame_suspects);
2772 if (sb->reverse && sb->contents_from)
2773 die(_("--contents and --reverse do not blend well."));
2775 if (!sb->repo)
2776 BUG("repo is NULL");
2778 if (!sb->reverse) {
2779 sb->final = find_single_final(sb->revs, &final_commit_name);
2780 sb->commits.compare = compare_commits_by_commit_date;
2781 } else {
2782 sb->final = find_single_initial(sb->revs, &final_commit_name);
2783 sb->commits.compare = compare_commits_by_reverse_commit_date;
2786 if (sb->final && sb->contents_from)
2787 die(_("cannot use --contents with final commit object name"));
2789 if (sb->reverse && sb->revs->first_parent_only)
2790 sb->revs->children.name = NULL;
2792 if (!sb->final) {
2794 * "--not A B -- path" without anything positive;
2795 * do not default to HEAD, but use the working tree
2796 * or "--contents".
2798 setup_work_tree();
2799 sb->final = fake_working_tree_commit(sb->repo,
2800 &sb->revs->diffopt,
2801 sb->path, sb->contents_from);
2802 add_pending_object(sb->revs, &(sb->final->object), ":");
2805 if (sb->reverse && sb->revs->first_parent_only) {
2806 final_commit = find_single_final(sb->revs, NULL);
2807 if (!final_commit)
2808 die(_("--reverse and --first-parent together require specified latest commit"));
2812 * If we have bottom, this will mark the ancestors of the
2813 * bottom commits we would reach while traversing as
2814 * uninteresting.
2816 if (prepare_revision_walk(sb->revs))
2817 die(_("revision walk setup failed"));
2819 if (sb->reverse && sb->revs->first_parent_only) {
2820 struct commit *c = final_commit;
2822 sb->revs->children.name = "children";
2823 while (c->parents &&
2824 !oideq(&c->object.oid, &sb->final->object.oid)) {
2825 struct commit_list *l = xcalloc(1, sizeof(*l));
2827 l->item = c;
2828 if (add_decoration(&sb->revs->children,
2829 &c->parents->item->object, l))
2830 BUG("not unique item in first-parent chain");
2831 c = c->parents->item;
2834 if (!oideq(&c->object.oid, &sb->final->object.oid))
2835 die(_("--reverse --first-parent together require range along first-parent chain"));
2838 if (is_null_oid(&sb->final->object.oid)) {
2839 o = get_blame_suspects(sb->final);
2840 sb->final_buf = xmemdupz(o->file.ptr, o->file.size);
2841 sb->final_buf_size = o->file.size;
2843 else {
2844 o = get_origin(sb->final, sb->path);
2845 if (fill_blob_sha1_and_mode(sb->repo, o))
2846 die(_("no such path %s in %s"), sb->path, final_commit_name);
2848 if (sb->revs->diffopt.flags.allow_textconv &&
2849 textconv_object(sb->repo, sb->path, o->mode, &o->blob_oid, 1, (char **) &sb->final_buf,
2850 &sb->final_buf_size))
2852 else
2853 sb->final_buf = read_object_file(&o->blob_oid, &type,
2854 &sb->final_buf_size);
2856 if (!sb->final_buf)
2857 die(_("cannot read blob %s for path %s"),
2858 oid_to_hex(&o->blob_oid),
2859 sb->path);
2861 sb->num_read_blob++;
2862 prepare_lines(sb);
2864 if (orig)
2865 *orig = o;
2867 free((char *)final_commit_name);
2872 struct blame_entry *blame_entry_prepend(struct blame_entry *head,
2873 long start, long end,
2874 struct blame_origin *o)
2876 struct blame_entry *new_head = xcalloc(1, sizeof(struct blame_entry));
2877 new_head->lno = start;
2878 new_head->num_lines = end - start;
2879 new_head->suspect = o;
2880 new_head->s_lno = start;
2881 new_head->next = head;
2882 blame_origin_incref(o);
2883 return new_head;
2886 void setup_blame_bloom_data(struct blame_scoreboard *sb)
2888 struct blame_bloom_data *bd;
2889 struct bloom_filter_settings *bs;
2891 if (!sb->repo->objects->commit_graph)
2892 return;
2894 bs = get_bloom_filter_settings(sb->repo);
2895 if (!bs)
2896 return;
2898 bd = xmalloc(sizeof(struct blame_bloom_data));
2900 bd->settings = bs;
2902 bd->alloc = 4;
2903 bd->nr = 0;
2904 ALLOC_ARRAY(bd->keys, bd->alloc);
2906 add_bloom_key(bd, sb->path);
2908 sb->bloom_data = bd;
2911 void cleanup_scoreboard(struct blame_scoreboard *sb)
2913 if (sb->bloom_data) {
2914 int i;
2915 for (i = 0; i < sb->bloom_data->nr; i++) {
2916 free(sb->bloom_data->keys[i]->hashes);
2917 free(sb->bloom_data->keys[i]);
2919 free(sb->bloom_data->keys);
2920 FREE_AND_NULL(sb->bloom_data);
2922 trace2_data_intmax("blame", sb->repo,
2923 "bloom/queries", bloom_count_queries);
2924 trace2_data_intmax("blame", sb->repo,
2925 "bloom/response-no", bloom_count_no);